Datasets:

Squadstack
/

conversational-streaming-asr-benchmark

sample_id int64	normalised_reference string	ref_words int32	arth_v1_raw string	arth_v1_normalised string	arth_v1_error_profile list	arth_v2_raw string	arth_v2_normalised string	arth_v2_error_profile list	nova_3_raw string	nova_3_normalised string	nova_3_error_profile list	scribe_v2_raw string	scribe_v2_normalised string	scribe_v2_error_profile list	saaras_v4_raw string	saaras_v4_normalised string	saaras_v4_error_profile list	soniox_v5_raw string	soniox_v5_normalised string	soniox_v5_error_profile list	flux_raw string	flux_normalised string	flux_error_profile list	cartesia_ink_2_raw string	cartesia_ink_2_normalised string	cartesia_ink_2_error_profile list	gemini_3_5_raw string	gemini_3_5_normalised string	gemini_3_5_error_profile list	chirp_3_raw string	chirp_3_normalised string	chirp_3_error_profile list	openai_gpt_live_raw string	openai_gpt_live_normalised string	openai_gpt_live_error_profile list
1	this call is now being recorded नहीं एक और number है दूसरा उनका है भाई घर पर नहीं है phone रखके गए है आप रात में करिए eight के बाद या seven के बाद हां ठीक ठीक	37	this call is now being recorded नहीं यह phone number है दूसरा उनका है घर पर नहीं है phone रखे गए आप रात रात में करिए eight के बाद या seven के बाद हां हां ठीक है	this call is now being recorded नहीं यह phone number है दूसरा उनका है घर पर नहीं है phone रखे गए आप रात रात में करिए eight के बाद या seven के बाद हां हां ठीक है	[ { "word": "एक → यह", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "और → phone", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "भाई", "op": "D", "turn_id": 2, "semantic_check":...	दिस कॉल इज नाउ बीइंग रिकॉर्डेड नहीं यह फोन नंबर है दूसरा उनका है घर पर नहीं है फोन रखे गए आप रात में करिए एक के बाद या सेवन के बाद हां हां ठीक है	this call is now being recorded नहीं यह phone number है दूसरा उनका है घर पर नहीं है phone रखे गए आप रात में करिए एक के बाद या seven के बाद हां हां ठीक है	[ { "word": "एक → यह", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "और → phone", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "भाई", "op": "D", "turn_id": 2, "semantic_check":...	this call is now being recorded नहीं यह phone number है दूसरा उनका है घर पर नहीं है phone रखकर गए आप रात में करिए eight के बाद या seven के बाद हां ठीक है	this call is now being recorded नहीं यह phone number है दूसरा उनका है घर पर नहीं है phone रखकर गए आप रात में करिए eight के बाद या seven के बाद हां ठीक है	[ { "word": "एक → यह", "op": "S", "turn_id": 2, "semantic_check": "Yes", "category": "number" }, { "word": "और → phone", "op": "S", "turn_id": 2, "semantic_check": "Yes", "category": "content" }, { "word": "भाई", "op": "D", "turn_id": 2, "semantic_check"...	This call is now being recorded. नहीं, phone number दूसरा उनका है, मैं घर पर नहीं हूँ। Phone रखे गए आपके। रात में करिए, के बाद में आप। Seven के बाद। हाँ, ठीक है।	this call is now being recorded नहीं phone number दूसरा उनका है मैं घर पर नहीं हूँ phone रखे गए आपके रात में करिए के बाद में आप seven के बाद हां ठीक है	[ { "word": "एक → phone", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "और", "op": "D", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "है", "op": "D", "turn_id": 2, "semantic_check": "No",...	This call is now being recorded. नहीं ये phone number है दूसरा उनका है घर पे नहीं है phone रखे गए आप। रात में करिए eight के बाद या। seven के बाद। हाँ ठीक ठीक।	this call is now being recorded नहीं ये phone number है दूसरा उनका है घर पर नहीं है phone रखे गए आप रात में करिए eight के बाद या seven के बाद हां ठीक ठीक	[ { "word": "एक → ये", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "और → phone", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "भाई", "op": "D", "turn_id": 2, "semantic_check":...	This call is now being recorded. नहीं, फोन नंबर दूसरा उनका है, घर पर नहीं है। फोन रख के गए आप। रात में करिए, 8 के बाद या। 7 के बाद। हाँ, ठीक।	this call is now being recorded नहीं phone number दूसरा उनका है घर पर नहीं है phone रखके गए आप रात में करिए eight के बाद या seven के बाद हां ठीक	[ { "word": "एक → phone", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "और", "op": "D", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "है", "op": "D", "turn_id": 2, "semantic_check": "No",...	This call is now being recorded. नहीं यह phone number है दूसरा है उनका है नहीं घर पर नहीं है phone रखे गए आप. रात में करी है eighth के बाद या Seven के बाद. हां.	this call is now being recorded नहीं यह phone number है दूसरा है उनका है नहीं घर पर नहीं है phone रखे गए आप रात में करी है eighth के बाद या seven के बाद हां	[ { "word": "एक → यह", "op": "S", "turn_id": 2, "semantic_check": "Yes", "category": "number" }, { "word": "और → phone", "op": "S", "turn_id": 2, "semantic_check": "Yes", "category": "content" }, { "word": "है", "op": "I", "turn_id": 2, "semantic_check":...	This call is now being recorded. नहीं, ये phone number दूसरा उनका है, माने, घर पे नहीं है phone रखे गए आप। रात में करिए, एक के बाद या seven के बाद। हाँ। ठीक, ठीक।	this call is now being recorded नहीं ये phone number दूसरा उनका है माने घर पर नहीं है phone रखे गए आप रात में करिए एक के बाद या seven के बाद हां ठीक ठीक	[ { "word": "एक → ये", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "और → phone", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "है", "op": "D", "turn_id": 2, "semantic_check": ...	This call is now being recorded. नहीं, ये फोन नंबर है दूसरा उनका है, घर पे नहीं है, फोन रख के गए हैं। रात में करिए 8:00 के बाद या। 7 के बाद हां, ठीक है।	this call is now being recorded नहीं ये phone number है दूसरा उनका है घर पर नहीं है phone रखके गए है रात में करिए eight के बाद या seven के बाद हां ठीक है	[ { "word": "एक → ये", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "और → phone", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "भाई", "op": "D", "turn_id": 2, "semantic_check":...	This call is now being recorded. नहीं है, फोन नंबर है दूसरा उनका है, घर पर नहीं है, फोन रख कर गया। रात में करिए 8:00 के बाद में आप। 7 ke baad	this call is now being recorded नहीं है phone number है दूसरा उनका है घर पर नहीं है phone रख कर गया रात में करिए eight के बाद में आप seven के बाद [MISSED]	[ { "word": "एक → है", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "और → phone", "op": "S", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "भाई", "op": "D", "turn_id": 2, "semantic_check":...	This call is now being recorded. नहीं फोन नंबर है दूसरा उनका है घर पे नहीं है फोन रखे गए आप रात्मक आठ के बाद या सेवन के बाद आ सही सही	this call is now being recorded नहीं phone number है दूसरा उनका है घर पर नहीं है phone रखे गए आप रात्मक eight के बाद या seven के बाद आ सही सही	[ { "word": "एक → phone", "op": "S", "turn_id": 2, "semantic_check": "Yes", "category": "number" }, { "word": "और", "op": "D", "turn_id": 2, "semantic_check": "No", "category": "content" }, { "word": "भाई", "op": "D", "turn_id": 2, "semantic_check": "No"...
2	"हां जी ***** available नहीं है\nहां हां हां हां ज(...TRUNCATED)	243	"हां जी ***** available नहीं है\nहां\nहां\nहां हां ज(...TRUNCATED)	"हां जी ***** available नहीं है\nहां हां हां हां ज(...TRUNCATED)	[{"word":"*** → ***","op":"S","turn_id":1,"semantic_check":"Yes","category":"name"},{"word":"(...TRUNCATED)	"हां जी ***** अवेलेबल नहीं है\nहां\nहां\nहां (...TRUNCATED)	"हां जी ***** available नहीं है\nहां हां हां हां ज(...TRUNCATED)	[{"word":"*****","op":"D","turn_id":1,"semantic_check":"Yes","category":"name"},{"word":"जी","op(...TRUNCATED)	"हां जी ***\n*** available नहीं है\nहां\nहां\nहां जी(...TRUNCATED)	"हां जी ***** available नहीं है\nहां हां हां जी बत(...TRUNCATED)	[{"word":"*** → **","op":"S","turn_id":1,"semantic_check":"Yes","category":"name"},{"word":"(...TRUNCATED)	"हाँ, जी, ***** available नहीं है।\nहाँ।\nहाँ।\nहाँ, (...TRUNCATED)	"हां जी ***** available नहीं है\nहां हां हां हां ज(...TRUNCATED)	[{"word":"*** → **","op":"S","turn_id":1,"semantic_check":"Yes","category":"name"},{"word":"(...TRUNCATED)	"जी ***** available नहीं है।\nहाँ।\nहाँ।\nहाँ हाँ।\n(...TRUNCATED)	"जी ***** available नहीं है\nहां हां हां हां जी बत(...TRUNCATED)	[{"word":"हां","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"***** (...TRUNCATED)	"हाँ जी, ***** अवेलेबल नहीं है।\nहाँ।\nहाँ।\n(...TRUNCATED)	"हां जी ***** available नहीं है\nहां हां हां हां ज(...TRUNCATED)	[{"word":"*** → **","op":"S","turn_id":1,"semantic_check":"Yes","category":"name"},{"word":"(...TRUNCATED)	"हां जी ***** available नहीं है.\nहां.\nहां.\nहां जी ब(...TRUNCATED)	"हां जी ***** available नहीं है\nहां हां हां जी बत(...TRUNCATED)	[{"word":"*** → **","op":"S","turn_id":1,"semantic_check":"Yes","category":"name"},{"word":"(...TRUNCATED)	"हाँ जी, जी ***** available नहीं हैं।\nहाँ।\nहाँ।\nह(...TRUNCATED)	"हां जी जी ***** available नहीं हैं\nहां हां हां ह(...TRUNCATED)	[{"word":"जी","op":"I","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"***** →(...TRUNCATED)	"जी ***** अवेलेबल नहीं है।\nहां।\nहां।\nहां (...TRUNCATED)	"जी ***** available नहीं है\nहां हां हां जी बताइए\ns(...TRUNCATED)	[{"word":"हां","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"***** (...TRUNCATED)	"हां जी, ***** अवेलेबल नहीं है।\nहां। हां। ह(...TRUNCATED)	"हां जी ***** available नहीं है\nहां हां हां जी बत(...TRUNCATED)	[{"word":"*** → **","op":"S","turn_id":1,"semantic_check":"Yes","category":"name"},{"word":"(...TRUNCATED)	"हाँ जी ***** अवेलेबल नहीं है\nहाँ\nहाँ\nहाँ (...TRUNCATED)	"हां जी ***** available नहीं है\nहां हां हां हां ज(...TRUNCATED)	[{"word":"*** → **","op":"S","turn_id":1,"semantic_check":"Yes","category":"name"},{"word":"(...TRUNCATED)
3	"hello जी\nहां आ रही है\nहां हां जी\nनहीं\norder म(...TRUNCATED)	25	"hello\nहां आ रही है\nहां हां जी\nनहीं मैं order म(...TRUNCATED)	"hello\nहां आ रही है\nहां हां जी\nनहीं\nमैं order (...TRUNCATED)	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"मैं(...TRUNCATED)	"हेलो\nहां आ रही है\nहां हां जी\nनहीं और म(...TRUNCATED)	"hello\nहां आ रही है\nहां हां जी\nनहीं\nऔर मेर(...TRUNCATED)	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"order →(...TRUNCATED)	"hello\nहां sorry\nहां\nहां जी\nनहीं\nमैं order मेरा\nor(...TRUNCATED)	"hello\nहां sorry\nहां हां जी\nनहीं\nमैं order मेरा\nord(...TRUNCATED)	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"आ → s(...TRUNCATED)	"Hello?\nहाँ, हाँ री।\nहाँ, हाँ जी।\nनहीं। Order म(...TRUNCATED)	"hello\nहां हां री\nहां हां जी\nनहीं\norder मेरा\n(...TRUNCATED)	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"आ → (...TRUNCATED)	"हाँ हाँ री।\nहाँ।\nहाँ जी।\nनहीं।\nऔर मेर(...TRUNCATED)	"[MISSED]\nहां हां री\nहां हां जी\nनहीं\nऔर मेरा(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"जी","o(...TRUNCATED)	"हेलो।\nहाँ, हाँ, आ रही है।\nहाँ। हाँ, जी।\n(...TRUNCATED)	"hello\nहां हां आ रही है\nहां हां जी\nनहीं\nऔर (...TRUNCATED)	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां(...TRUNCATED)	"Hello.\nSorry.\nहां हां जी.\nनहीं. मेरा order मेरा order (...TRUNCATED)	"hello\nsorry\nहां हां जी\nनहीं\nमेरा order\nमेरा order थ(...TRUNCATED)	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां(...TRUNCATED)	"Hello?\nहाँ, sorry.\nहाँ। हाँ जी।\nनहीं। और मेरा- (...TRUNCATED)	"hello\nहां sorry\nहां हां जी\nनहीं\nऔर मेरा\nहां (...TRUNCATED)	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"आ → s(...TRUNCATED)	"Hello.\nAa rahi.\nहां जी।\nथा वो मेरे को रिटर्न\nReturn(...TRUNCATED)	"hello\nआ रही\nहां जी\n[MISSED]\n[MISSED]\nथा वो मेरे को retur(...TRUNCATED)	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां(...TRUNCATED)	"हेलो।\nहां, आ रही। हां, हां जी।\nनहीं।\nऔ(...TRUNCATED)	"hello\nहां आ रही\nहां हां जी\nनहीं\nऔर मेरा\nम(...TRUNCATED)	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"है","(...TRUNCATED)	"हैलो\nSorry\nहाँ हाँ जी\nनहीं और मेरा ऑर्डर (...TRUNCATED)	"hello\nsorry\nहां हां जी\nनहीं\nऔर मेरा\norder था वने(...TRUNCATED)	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां(...TRUNCATED)
4	hello जी जी जी madam क्या पूछ क्या	8	hello जी जी जी madam क्या	hello जी जी जी madam क्या	[{"word":"पूछ","op":"D","turn_id":3,"semantic_check":"No","category":"content"},{"word":"क(...TRUNCATED)	हेलो जी जी मैडम क्या	hello जी जी [MISSED] madam क्या	[{"word":"जी","op":"D","turn_id":2,"semantic_check":"No","category":"filler"},{"word":"पूछ(...TRUNCATED)	hello जी	hello जी [MISSED] [MISSED]	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"जी","(...TRUNCATED)	जी। जी।	जी जी [MISSED]	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"जी","o(...TRUNCATED)	जी। Madam. क्या हुआ? तो।	जी [MISSED] madam क्या हुआ तो	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"जी","o(...TRUNCATED)	रुको। जी। मैडम, क्या—	रुको जी madam क्या	[{"word":"hello → रुको","op":"S","turn_id":1,"semantic_check":"Yes","category":"content"},(...TRUNCATED)	Hello. जी जी. क्या So Okay. Okay. Okay.	hello जी जी क्या so okay okay okay	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"क्य(...TRUNCATED)	जी, जी। जी। माइक ऑन है, क्या हुआ?	जी जी जी माइक ऑन है क्या हुआ	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"madam → (...TRUNCATED)	जी। जी। क्या	जी जी क्या	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"जी","o(...TRUNCATED)	Hello. जी जी। क्या हो?	hello जी जी क्या हो	[{"word":"जी","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"madam","o(...TRUNCATED)	Hello जी जी Yes जी क्या	hello जी जी yes जी क्या	[{"word":"yes","op":"I","turn_id":2,"semantic_check":"No","category":"filler"},{"word":"madam","op":(...TRUNCATED)
5	"हां\nhello\nदो सौ piece दो सौ piece\nहां\nblack black black blue\nह(...TRUNCATED)	15	"नौ number नौ\nजी maam\nकौन बोल रहा है\nhello\nहां\nबो(...TRUNCATED)	"नौ number नौ जी maam कौन बोल रहा है\nhello\nहां बोल(...TRUNCATED)	[{"word":"नौ","op":"I","turn_id":1,"semantic_check":"Yes","category":"number"},{"word":"number",(...TRUNCATED)	"टुमॉरो *****\nकौन सा बात है\nहेलो\nहां\nदो\nब(...TRUNCATED)	"टुमॉरो नंबर कौन सा बात है\nhello\nहां दो\n[MISSED(...TRUNCATED)	[{"word":"टुमॉरो","op":"I","turn_id":1,"semantic_check":"Yes","category":"content"},{"wo(...TRUNCATED)	hello black black black blue हां हां fifteen	[MISSED] hello [MISSED] [MISSED] black black black blue हां हां fifteen	[{"word":"हां","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"दो(...TRUNCATED)	"मालूम नहीं, मैं नौ।\nअरे, तो अभी तो बात क(...TRUNCATED)	"मालूम नहीं मैं नौ अरे तो अभी तो बात कर (...TRUNCATED)	[{"word":"मालूम","op":"I","turn_id":1,"semantic_check":"Yes","category":"content"},{"word"(...TRUNCATED)	"नानो नमो।\nकैसे हम लोग\nहाँ।\nहाँ।\nजो बो(...TRUNCATED)	"नानो नमो कैसे हम लोग हां हां जो बोलेंग(...TRUNCATED)	[{"word":"नानो","op":"I","turn_id":1,"semantic_check":"Yes","category":"content"},{"word":"(...TRUNCATED)	"नाम और ***** देना होगा।\nनाम-नंबर कैसे?\nतो (...TRUNCATED)	"नाम और नंबर देना होगा नाम नंबर कैसे तो (...TRUNCATED)	[{"word":"नाम","op":"I","turn_id":1,"semantic_check":"Yes","category":"content"},{"word":"औ(...TRUNCATED)	Hello. Two Black black black Fifteen.	[MISSED] hello दो [MISSED] black black black fifteen	[{"word":"हां","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"सौ(...TRUNCATED)	"nano numbers and all.\nबात कर रहे थे, कार मामला थोड़ा(...TRUNCATED)	"nano numbers and all बात कर रहे थे कार मामला थोड़ा (...TRUNCATED)	[{"word":"nano","op":"I","turn_id":1,"semantic_check":"Yes","category":"content"},{"word":"numbers",(...TRUNCATED)	"नानु ***** देना\nहां।\nबहुत-बहुत लोग बोलें(...TRUNCATED)	"नानु नंबर देना हां\nबहुत बहुत लोग बोले(...TRUNCATED)	[{"word":"नानु","op":"I","turn_id":1,"semantic_check":"Yes","category":"name"},{"word":"न(...TRUNCATED)	*****	[MISSED] [MISSED] ***** [MISSED] [MISSED] [MISSED]	[{"word":"हां","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"hello"(...TRUNCATED)	"नानु ನಮ್ಮನ್ನು\nneden\nबात\n嗯\nबहुत लोग बोले\n(...TRUNCATED)	"नानु ನಮ್ಮನ್ನು neden बात 嗯 बहुत लोग बोले क(...TRUNCATED)	[{"word":"नानु","op":"I","turn_id":1,"semantic_check":"Yes","category":"content"},{"word":"(...TRUNCATED)
6	"hello\nहां\nहां\nजी हां बोलिए\nhello\nhello\nnumber का छो(...TRUNCATED)	76	"जी हां बोलिए yes\nok\nhello\nhello\nnumber का छोड़ो many curing (...TRUNCATED)	"[MISSED]\n[MISSED]\n[MISSED]\nजी हां बोलिए yes ok\nhello\nhello\nnumber का (...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां"(...TRUNCATED)	"जी हां बोलिए\nहेलो\nहेलो\nनंबर का छोडो म(...TRUNCATED)	"[MISSED]\n[MISSED]\n[MISSED]\nजी हां बोलिए\nhello\nhello\nnumber का छो(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां"(...TRUNCATED)	"का छोड़ो वह main ही query है मेरे पास\nउसकी\nquesti(...TRUNCATED)	"[MISSED]\n[MISSED]\n[MISSED]\n[MISSED]\n[MISSED]\n[MISSED]\nका छोड़ो वह main ही (...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां"(...TRUNCATED)	"हाँ।\nजी हाँ, बोलिए।\nनागराज।\nHello, ma'am.\nHello. (...TRUNCATED)	"[MISSED]\n[MISSED]\nहां\nजी हां बोलिए नागराज\nhello ma am\nhe(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां"(...TRUNCATED)	"हाँ।\nक्या बोलिए?\nदेखाना तो ठीक।\nना दे(...TRUNCATED)	"[MISSED]\nहां\n[MISSED]\nक्या बोलिए\nदेखाना तो ठीक (...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां"(...TRUNCATED)	"हाँ।\nजी, हाँ, बोलिए।\nहेलो।\nहेलो। नंबर (...TRUNCATED)	"[MISSED]\nहां\n[MISSED]\nजी हां बोलिए\nhello\nhello\nnumber का छो(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां"(...TRUNCATED)	"हां. Okay.\nHello\nHello. नमस्कार छोड़ो, manic query है मेर(...TRUNCATED)	"[MISSED]\nहां\n[MISSED]\nokay\nhello\nhello\nनमस्कार छोड़ो manic quer(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां"(...TRUNCATED)	"Hello.\nHello नंबर तीन का छोड़ो, many query है मेरे पा(...TRUNCATED)	"hello\n[MISSED]\n[MISSED]\n[MISSED]\n[MISSED]\nhello\nnumber तीन का छोड़ो many que(...TRUNCATED)	[{"word":"हां","op":"D","turn_id":2,"semantic_check":"No","category":"filler"},{"word":"हा(...TRUNCATED)	"जी हां, बोलिए।\nHello\nNumber ka chhoro, main inquiry hai mere pass.\nउस(...TRUNCATED)	"[MISSED]\n[MISSED]\n[MISSED]\nजी हां बोलिए\nhello\n[MISSED]\nnumber का छ(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां"(...TRUNCATED)	"हां, बोलिए।\nइतना वक्त ही ना देना\nहेलो(...TRUNCATED)	"[MISSED]\n[MISSED]\n[MISSED]\nहां बोलिए\nhello\nhello\nnumber का छोड़ो m(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां"(...TRUNCATED)	"हाँ\nजी हाँ बोलिए\nन\nHello\nHello আমার মেনি\nउसक(...TRUNCATED)	"[MISSED]\nहां\n[MISSED]\nजी हां बोलिए\nन hello\nhello\nআমার ম(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां"(...TRUNCATED)
7	"हां hello\nहां हां ma'am\nहां हां ma'am\nदस से बारह\n(...TRUNCATED)	20	"हां\nhello\nहां\nmaam\nदस से बारह\nदस\nखुद ही deliver क(...TRUNCATED)	"हां hello\nहां maam\n[MISSED]\nदस से बारह\nदस\nखुद ही del(...TRUNCATED)	[{"word":"हां → maam","op":"S","turn_id":2,"semantic_check":"No","category":"filler"},{"word(...TRUNCATED)	"हां\nहेलो\nहां मैम\nदस से बारह\nदस\nखुद ही (...TRUNCATED)	"हां hello\nहां ma'am\n[MISSED]\nदस से बारह\nदस\nखुद ही de(...TRUNCATED)	[{"word":"हां","op":"D","turn_id":2,"semantic_check":"No","category":"filler"},{"word":"हा(...TRUNCATED)	"हां\nhello\nहां\nma'am\nma'am\nदस से बारह\nदस\nही deliver क(...TRUNCATED)	"हां hello\nहां ma'am\nma'am\nदस से बारह\nदस\nही deliver कर(...TRUNCATED)	[{"word":"हां","op":"D","turn_id":2,"semantic_check":"No","category":"filler"},{"word":"हा(...TRUNCATED)	"हाँ।\nहाँ, नो।\nहाँ। हाँ, ma'am।\nHello?\nHello?\nMa'am.\nह(...TRUNCATED)	"हां हां नो\nहां हां ma'am\nhello hello ma'am हां बोल रह(...TRUNCATED)	[{"word":"हां","op":"I","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"hello (...TRUNCATED)	"सही सही।\nहाँ।\nहाँ Ma'am.\nहं।\nहाँ Ma'am.\nदस से (...TRUNCATED)	"सही सही\nहां हां ma'am\nहं हां ma'am\nदस से बारह\(...TRUNCATED)	[{"word":"हां → सही","op":"S","turn_id":1,"semantic_check":"No","category":"filler"},{(...TRUNCATED)	"हाँ?\nहेलो?\nहाँ, मैम।\nहाँ, मैम।\n10 से 12।\n10?\n(...TRUNCATED)	"हां hello\nहां ma'am\nहां ma'am\nदस से बारह\nदस\nखुली(...TRUNCATED)	[{"word":"हां","op":"D","turn_id":2,"semantic_check":"No","category":"filler"},{"word":"हा(...TRUNCATED)	"हां?\nHello.\nma'am.\nma'am.\nवाला.\nदस.\nखुद ही deliver करते (...TRUNCATED)	"हां hello\nma'am\nma'am\nवाला\nदस\nखुद ही deliver करतें ह(...TRUNCATED)	[{"word":"हां","op":"D","turn_id":2,"semantic_check":"No","category":"filler"},{"word":"हा(...TRUNCATED)	"Huh?\nHello?\nUm, ma'am.\nUm, ma'am.\nदस्ते वाला।\nदस।\nतो delivery d(...TRUNCATED)	"huh hello\num ma'am\num ma'am\nदस्ते वाला\nदस\nतो delivery deliver कर(...TRUNCATED)	[{"word":"हां → huh","op":"S","turn_id":1,"semantic_check":"No","category":"filler"},{"word"(...TRUNCATED)	"हां।\nहेलो।\nहां मैम।\nHmm.\nI'm\nKasturba\n10\nखुद ही (...TRUNCATED)	"हां hello\nहां ma'am\nhmm im\nkasturba\nदस\nखुद ही deliver करते(...TRUNCATED)	[{"word":"हां","op":"D","turn_id":2,"semantic_check":"No","category":"filler"},{"word":"हा(...TRUNCATED)	हेलो। 10 खुद ही डिलीवर करते। Kya kar rahe ho?	"hello\n[MISSED]\n[MISSED]\n[MISSED]\nदस\nखुद ही deliver करतें\nkya kar rahe(...TRUNCATED)	[{"word":"हां","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हा(...TRUNCATED)	"हाँ\nHello\nमैम\n嗯\nI'm\n嗯\nउसके\nदस\nOkay डिलीवर\nJuly\n(...TRUNCATED)	हां hello ma'am 嗯 i'm 嗯 उसके दस okay deliver july न	[{"word":"हां","op":"D","turn_id":2,"semantic_check":"No","category":"filler"},{"word":"हा(...TRUNCATED)
8	"कौन *****\nहां हां हां बोलिए\nसुनो maam\nनहीं maa(...TRUNCATED)	258	"कौन हमारी company\nहां हां हां बोलिए\nनहीं maam(...TRUNCATED)	"कौन हमारी company\nहां हां हां बोलिए\n[MISSED]\nनह(...TRUNCATED)	[{"word":"*** → **","op":"S","turn_id":1,"semantic_check":"Yes","category":"name"},{"word":"(...TRUNCATED)	"हां हां हां बोलिए\nनहीं मैम\nमैंने जाने(...TRUNCATED)	"[MISSED]\nहां हां हां बोलिए\n[MISSED]\nनहीं maam मैंन(...TRUNCATED)	[{"word":"कौन","op":"D","turn_id":1,"semantic_check":"Yes","category":"content"},{"word":"****(...TRUNCATED)	"हां\nहां हां बोलिए\nनहीं ma'am\nमैंने जाने (...TRUNCATED)	"[MISSED]\nहां हां हां बोलिए\n[MISSED]\nनहीं maam मैंन(...TRUNCATED)	[{"word":"कौन","op":"D","turn_id":1,"semantic_check":"Yes","category":"content"},{"word":"****(...TRUNCATED)	"200 अधिक हैं, मिले।\nहाँ, हाँ, हाँ बोलिए।\n(...TRUNCATED)	"200 अधिक हैं मिले\nहां हां हां बोलिए\nबोल(...TRUNCATED)	[{"word":"200","op":"I","turn_id":1,"semantic_check":"Yes","category":"number"},{"word":"कौन (...TRUNCATED)	"*****\nहाँ हाँ हाँ बोलिए।\nदुरुंग\nहम्म हम्(...TRUNCATED)	"*****\nहां हां हां बोलिए\nदुरुंग हम्म हम्म\(...TRUNCATED)	[{"word":"*** → ***","op":"S","turn_id":1,"semantic_check":"Yes","category":"content"},{"word"(...TRUNCATED)	"दोनों। अभी *****\nहाँ। हाँ, हाँ, बोलिए।\nकर(...TRUNCATED)	"दोनों अभी *****\nहां हां हां बोलिए\nकरूँगा\(...TRUNCATED)	[{"word":"कौन → दोनों","op":"S","turn_id":1,"semantic_check":"Yes","category":"con(...TRUNCATED)	"हां हां हां बोलिए.\nHello, ma'am.\nउसमें. नहीं ma'am\(...TRUNCATED)	"[MISSED]\nहां हां हां बोलिए\nhello maam\nउसमें नहीं m(...TRUNCATED)	[{"word":"कौन","op":"D","turn_id":1,"semantic_check":"Yes","category":"content"},{"word":"****(...TRUNCATED)	"नमो *****\nहाँ, हाँ, हाँ, बोलिए।\nसुनो ना।\nHmm, (...TRUNCATED)	"नमो *****\nहां हां हां बोलिए\nसुनो ना hmm hmm hmm hmm(...TRUNCATED)	[{"word":"कौन → नमो","op":"S","turn_id":1,"semantic_check":"Yes","category":"content"}(...TRUNCATED)	"*****\nहां हां हां बोलिए।\nकरूंगा।\nHmm.\nनहीं (...TRUNCATED)	"*****\nहां हां हां बोलिए\nकरूंगा hmm\nनहीं maam tha(...TRUNCATED)	[{"word":"कौन","op":"D","turn_id":1,"semantic_check":"Yes","category":"content"},{"word":"स(...TRUNCATED)	"कौन है *****\nहां, हां, हां, बोलिए।\nकरूंगा। (...TRUNCATED)	"कौन है *****\nहां हां हां बोलिए\nकरूंगा\nनह(...TRUNCATED)	[{"word":"है","op":"I","turn_id":1,"semantic_check":"No","category":"content"},{"word":"***** (...TRUNCATED)	"तो ***\n*** बोलिए\nसुनो\nहम्म हम्म\nहम्म नह(...TRUNCATED)	"तो *****\nहां हां हां बोलिए\nसुनो हम्म हम्म(...TRUNCATED)	[{"word":"कौन → तो","op":"S","turn_id":1,"semantic_check":"Yes","category":"content"},{"(...TRUNCATED)
9	"hello\nहां बोलिए hello hello बात कीजिए hello recording आहे\n(...TRUNCATED)	43	"हां बोलिए बोलिए\nhello\nhello बात कीजिए\nyes sir hello\nr(...TRUNCATED)	"[MISSED]\nहां बोलिए बोलिए hello hello बात कीजिए yes sir h(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"बोल(...TRUNCATED)	"हां बोलिए बोलिए\nहेलो\nहेलो बात कीजिए\n(...TRUNCATED)	"[MISSED]\nहां बोलिए बोलिए hello hello बात कीजिए क्य(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"बोल(...TRUNCATED)	"हां बोलिए\nhello\nबात कीजिए\nhello\nrecording i have\nहां ब(...TRUNCATED)	"[MISSED]\nहां बोलिए hello बात कीजिए hello recording i have\n[MISSED(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"hello","op(...TRUNCATED)	"हाँ, बोलिए।\nHello.\nHello, बात कीजिए।\nHello.\nहम्म, (...TRUNCATED)	"[MISSED]\nहां बोलिए hello hello बात कीजिए hello हम्म हम(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हम्(...TRUNCATED)	"हाँ बोलिए।\nHello.\nHello बात कीजिए।\nक्या?\nHello.\n(...TRUNCATED)	"[MISSED]\nहां बोलिए hello hello बात कीजिए क्या hello हम(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"क्य(...TRUNCATED)	"नम बोलिए।\nहेलो।\nहेलो, बात कीजिए।\nहेल(...TRUNCATED)	"[MISSED]\nनम बोलिए hello hello बात कीजिए hello recording आ रह(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां (...TRUNCATED)	"हां बोलिए.\nHello.\nHello बात कीजिए.\nHello.\nवह. Recording आ(...TRUNCATED)	"[MISSED]\nहां बोलिए hello hello बात कीजिए hello वह recording (...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"वह","o(...TRUNCATED)	"That's all हाँ, बोलिए।\nHello!\nHello! बात कीजिए।\nHello!\nHmm,(...TRUNCATED)	"that's all\nहां बोलिए hello hello बात कीजिए hello hmm hmm hmm recor(...TRUNCATED)	[{"word":"that's","op":"I","turn_id":1,"semantic_check":"Yes","category":"content"},{"word":"hello (...TRUNCATED)	"हां बोलिए।\nHello.\nHello, baat kijiye.\nHello.\nhai\nSee *****\nPlease let me kn(...TRUNCATED)	"[MISSED]\nहां बोलिए hello hello बात कीजिए hello है\nsee ***** p(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"recording (...TRUNCATED)	"नाम बोलिए।\nहेलो।\nहेलो, बात कीजिए।\nहे(...TRUNCATED)	"[MISSED]\nनाम बोलिए hello hello बात कीजिए hello recording आहे(...TRUNCATED)	[{"word":"hello","op":"D","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"हां (...TRUNCATED)	"Okay हाँ बोलिए\nHello\nHello बात कीजिए।\nHello\nOh.\nरेकॉ(...TRUNCATED)	"okay\nहां बोलिए hello hello बात कीजिए hello oh recording आहे\(...TRUNCATED)	[{"word":"hello → okay","op":"S","turn_id":1,"semantic_check":"No","category":"filler"},{"word":"o(...TRUNCATED)
10	"हां बोलिए\nmetal चाहिए\nmet\nTV panel\nTV panel चाहिए हमे(...TRUNCATED)	12	"हां बोलिए\nmetal चाहिए\nमैं\ndv panel\ndg panel चाहिए ह(...TRUNCATED)	"हां बोलिए\nmetal चाहिए\nमैं\nDV panel\nDG panel चाहिए ह(...TRUNCATED)	[{"word":"met → मैं","op":"S","turn_id":3,"semantic_check":"Yes","category":"content"},{"wor(...TRUNCATED)	"हां बोलिए\nमैसेज चाहिए\nमैं\nडी वी पैनल\n(...TRUNCATED)	"हां बोलिए\nमैसेज चाहिए\nमैं\nDV panel\npanel चाहि(...TRUNCATED)	[{"word":"metal → मैसेज","op":"S","turn_id":2,"semantic_check":"Yes","category":"content(...TRUNCATED)	"हां बोलिए\nmetal चाहिए\nmetal\ntv panel\nजी जी panel चाहि(...TRUNCATED)	"हां बोलिए\nmetal चाहिए\nmetal\nTV panel\nजी जी panel चाहि(...TRUNCATED)	[{"word":"met → metal","op":"S","turn_id":3,"semantic_check":"Yes","category":"content"},{"word":"(...TRUNCATED)	"हाँ, बोलिए।\nमैं कल से ही है।\nमैं...\nDD panel ह(...TRUNCATED)	"हां बोलिए\nमैं कल से ही है\nमैं\nDD panel है\nDD pa(...TRUNCATED)	[{"word":"मैं","op":"I","turn_id":2,"semantic_check":"Yes","category":"content"},{"word":"क(...TRUNCATED)	"हाँ बोलिए।\nmuscle चाहिए।\nमैं।\nबीबी panel।\nDG pa(...TRUNCATED)	"हां बोलिए\nmuscle चाहिए\nमैं\nबीबी panel\nDG panel चा(...TRUNCATED)	[{"word":"metal → muscle","op":"S","turn_id":2,"semantic_check":"Yes","category":"content"},{"word(...TRUNCATED)	"हाँ, बोलिए।\nमेटल चाहिए।\nमैं।\nडीवी पै(...TRUNCATED)	"हां बोलिए\nmetal चाहिए\nमैं\nडीवी panel\nडीवी panel(...TRUNCATED)	[{"word":"met → मैं","op":"S","turn_id":3,"semantic_check":"Yes","category":"content"},{"wor(...TRUNCATED)	"हां बोलिए.\nMetal चाहिए.\nमैं\nTV panel.\nEasy panel चाहि(...TRUNCATED)	"हां बोलिए\nmetal चाहिए\nमैं\nTV panel\neasy panel चाहिए (...TRUNCATED)	[{"word":"met → मैं","op":"S","turn_id":3,"semantic_check":"Yes","category":"content"},{"wor(...TRUNCATED)	"हाँ बोलिए।\nमटर चाहिए?\nमैं-\nTV panel\nTV panel चाह(...TRUNCATED)	"हां बोलिए\nमटर चाहिए\nमैं\nTV panel\nTV panel चाहिए(...TRUNCATED)	[{"word":"metal → मटर","op":"S","turn_id":2,"semantic_check":"Yes","category":"content"},{"w(...TRUNCATED)	"हां, बोलिए।\nMetal chahiye.\nबीवी पैनल\nडीजी पैनल(...TRUNCATED)	"हां बोलिए\nmetal चाहिए\n[MISSED]\nबीवी panel\nडीजी panel (...TRUNCATED)	[{"word":"met","op":"D","turn_id":3,"semantic_check":"Yes","category":"content"},{"word":"TV → ब(...TRUNCATED)	"क्या बोलिए?\nमक्खन चाहिए।\nमैं\nTV panel\nTV panel ch(...TRUNCATED)	"क्या बोलिए\nमक्खन चाहिए\nमैं\nTV panel\nTV panel चा(...TRUNCATED)	[{"word":"हां → क्या","op":"S","turn_id":1,"semantic_check":"Yes","category":"conten(...TRUNCATED)	"हाँ बोलिए\nما\nमेट\nटीवी पैनल\nटीवी पैनल (...TRUNCATED)	हां बोलिए ما met TV panel TV panel चाहिए हमे [MISSED]	[{"word":"metal → ما","op":"S","turn_id":2,"semantic_check":"Yes","category":"content"},{"word":(...TRUNCATED)

End of preview. Expand in Data Studio

At a glance: 863 real telesales calls, 11 ASR systems, 5 business domains, a 3x3 grid of code-mixing and noise, 8 kHz telephony.

SquadStack Conversational Streaming ASR Benchmark (8 kHz)

Version 1.0.0 · maintained by SquadStack

Schema · Leaderboard · Latency · Submit a system · Licence · Terms of use

Key takeaways

What this is. 863 real Hindi–English telesales calls (5.53 hours of customer speech, 8 kHz phone audio), human-transcribed turn by turn, and 11 speech recognisers scored on them. The question it answers: which recogniser should run inside an Indian voice agent, judged on what the customer meant and how fast the transcript arrives?

What to expect from it.

Raw WER overstates the problem. It is 27–55% across the 11 systems. Semantic WER, which counts only errors that would change what the agent does, is 9–22%. For the median system, 32% of word errors matter.
The top four are close. Soniox v5 (8.77%), Arth V2 (8.85%), Arth V1 (9.17%) and Nova 3 (9.31%) sit within about half a point of each other.
Noise costs more than code-mixing. Clean to noisy calls adds 4–12 points; low to high code-mixing adds 2–5.

Semantic WER

Semantic WER, lower is better, pooled over about 67,000 reference words; ★ marks SquadStack's in-house systems. Solid bar = errors that change meaning; pale bar = all word errors (raw WER). Full tables in LEADERBOARD.md.

Latency against accuracy

80th-percentile final-transcript latency (FTR P80, ms) against semantic WER for ten systems on the locked 50-call set.

Semantic WER against FTR P80: the 80th percentile, over all turns, of the time from our end-of-turn finalize signal to the vendor's final reply for that turn. Latency is from the locked 50-call set streamed in real time from a Delhi server; semantic WER is from the 863-call run. Chirp 3 does not support a client-side finalize, so it has no FTR. Method in LATENCY.md, table in LEADERBOARD.md.

It is an evaluation set. It is not for training, and training on it is a licence violation.

What this measures — and what it does not

It measures	It does not measure
Transcription of spontaneous Hinglish over 8 kHz telephony	Batch transcription quality on clean, wide-band audio
Behaviour on unsegmented audio that is about 80% non-speech	Per-utterance accuracy on pre-segmented clips
Whether errors change what a downstream agent would do	Literal string fidelity as an end in itself
80th-percentile time from our end-of-turn signal to the final transcript (50-call set)	Throughput, cost at scale, or GPU efficiency
Robustness across five Indian consumer-sales verticals	Any non-telesales domain: medical, legal, broadcast, meetings
Hindi and English in a code-mixed register	The other 20 scheduled Indian languages
Customer (lead-side) audio only	Agent-side audio, or full-duplex diarization

If your question is "which streaming STT should run inside my Indian voice agent", this benchmark is built for you. If your question is "which ASR model is best", it is not: use it alongside the Open ASR Leaderboard, FLEURS and IndicVoices, not instead of them.

At a glance


Calls	863
Audio	29.58 h total · 5.53 h of customer speech (18.7%)
Format	8 kHz, mono, PCM16, customer channel extracted from stereo telephony
Language	Hindi–English code-mixed: 16% low, 53% mid, 31% high code-mixing
Domains	Marketplace, Logistics, Travel, BFSI, Education
Campaigns	19
Gender	16.9% female / 83.1% male (human-annotated)
References	Human-annotated, turn-separated, code-switched (Hindi in Devanagari, English in Latin)
Split	`test` only. There is no train split, by design.
Systems benchmarked	11: SquadStack in-house and commercial APIs
Headline metric	Semantic WER (pooled), with raw WER, its edit types, error classes and P80 latency alongside

Dataset structure

Two tables, both 863 rows, one row per call, joined on sample_id. Loading the dataset without a config name gives you test.

from datasets import load_dataset

bench = load_dataset("Squadstack/conversational-streaming-asr-benchmark", "test")["test"]
res   = load_dataset("Squadstack/conversational-streaming-asr-benchmark", "results")["test"]

Config	Rows	Columns	What it is
`test`	863	11	the benchmark: audio, both versions of the reference, and the call's acoustic and demographic profile
`results`	863	37	the same audio, what the 11 systems returned and every error each one made, as a structured per-error profile

The split is deliberate. test is the thing being released and does not change when a system is added or re-run; results is a snapshot of eleven systems and will. A new baseline never touches the benchmark itself.

Recomputing our numbers

Every accuracy figure in this card is a count over the error profiles in results, and both tables ship:

prof = res["soniox_v5_error_profile"]              # one list per call
N    = sum(res["ref_words"])

errs = [e for call in prof for e in call]
sem  = [e for e in errs if e["semantic_check"] == "Yes"]

raw_wer      = len(errs) / N * 100
semantic_wer = len(sem)  / N * 100

If your numbers disagree with ours, we treat it as our bug and log it as an erratum. Latency is measured while audio streams and comes from the run record, which the maintainers hold.

Fields

test: 11 fields, one row per call

Field	Type	Description
`sample_id`	int	Serial number, 1 to 863. Stable: cite this, never a row index. Joins to `results`.
`audio`	Audio(8 kHz)	Customer-channel clip, unsegmented: ringing, hold and silence retained. Personal information is replaced by a 1 kHz beep.
`duration_sec`	float32	Whole call, silence included
`speech_duration_sec`	float32	Customer speech only (voice-activity detection)
`business_domain`	string	`BFSI` / `Travel` / `Marketplace` / `Logistics` / `Education`
`speaker_gender`	string	`male` / `female`, human-annotated
`noise_band`	string	`clean` / `moderate` / `noisy`, from DNSMOS `bak` and NISQA `noi` on identical 9-second pure-speech chunks
`cmi_band`	string	`low_mix` / `mid_mix` / `high_mix`: code-mixing index of the normalised human reference, banded at 0.15 and 0.35
`original_reference`	string	Human transcript before normalisation, turn-separated. Shipped so you can disagree with our normalisation and redo it from source. Personal information is masked as `*****`.
`normalised_reference`	string	The reference every system is scored against. Cleaned once, then locked; one line per turn. Personal information is masked as `*****`.
`ref_words`	int32	Words in `normalised_reference`, counted before masking

results: 37 fields, one row per call

Same 863 rows, joined on sample_id. Three columns per system, so adding a twelfth system adds three columns and changes nothing else.

Field	Type	Description
`sample_id`	int	Join key to `test`
`audio`	Audio(8 kHz)	The same redacted clip as in `test`, repeated so the viewer plays it next to the transcripts
`normalised_reference`	string	Repeated from `test` so this table is scoreable on its own
`ref_words`	int32	Reference word count, counted before masking
`<system>_raw`	string	What the vendor returned, with its own punctuation and digits; personal information is masked as `*****`
`<system>_normalised`	string	After normalisation and alignment to the reference turns, one line per reference turn (an empty line where the system returned nothing for that turn); the text that was scored, with personal information masked as `*****`
`<system>_error_profile`	list<struct>	Every error that system made on that call, one record each; words involved in masked text appear as `*****`

Each error record: word (बीस → बिस for a substitution, the token alone for a deletion or insertion), op (S · D · I), turn_id, semantic_check (Yes · No: would this change what the agent does next), and category (number · negation · name · commitment · content · filler · spelling).

The 11 systems, in column order

Prefix	System	Vendor
`arth_v1`	Arth V1	SquadStack
`arth_v2`	Arth V2	SquadStack
`nova_3`	Nova 3	Deepgram
`scribe_v2`	Scribe v2 Realtime	ElevenLabs
`saaras_v4`	Saaras v4	Sarvam
`soniox_v5`	Soniox v5	Soniox
`flux`	Flux	Deepgram
`cartesia_ink_2`	Cartesia ink-2	Cartesia
`gemini_3_5`	Gemini 3.5 Transcribe	Google
`chirp_3`	Chirp 3	Google
`openai_gpt_live`	OpenAI gpt-live-transcribe	OpenAI

Redaction

Personal information (the lead's identity and SquadStack's customer names) is removed everywhere. In the audio it is a 1 kHz beep. In every text field (original_reference, normalised_reference, the system _raw and _normalised texts and the word of each error) it is masked as *****, and the same words are masked everywhere within a call. A mask can also cover a word next to the personal information.

The numbers in this card were computed before masking, and every error record is kept, so counting the error profiles reproduces them exactly. Scoring the masked texts yourself gives slightly different values: the masked words match each other and the reference has about 900 fewer words, so each system's raw and semantic WER come out about 0.3 to 0.5 points lower, with the same ranking. call_sid and any other call identifier are not published; sample_id is a serial number.

How the 863 were chosen

Zero contamination with Arth's training data. Every one of the 863 calls was checked against the audio used to pre-train and fine-tune SquadStack's Arth models. No recording and no speaker in this benchmark appears in either training set.

Every stage is a fixed rule, not a hand pick: positive-outcome calls only (the customer engaged), calls profiled for speech share, noise, code-mixing, vocabulary diversity and content, and a selection that fills a 3 × 3 grid of code-mixing against noise with no domain above a quarter of the speech hours. Personal information and client names are removed from the audio and the transcripts.

Diversity

A benchmark that concentrates on one or two verticals cannot tell you whether a system will hold up on a vertical you have not tested, so the selection is spread deliberately across code-mixing density, acoustic condition and business vertical. Gender is reported, not balanced.

Panels: code-mixing by noise grid, domain mix, call length, customer speech per call, speaking rate and vocabulary.

Stratification: code-mixing × acoustic condition

Nine cells, all populated, from 9 calls (low-mix clean) to 265 (mid-mix moderate). The grid is unbalanced, and the marginal slices inherit that. Read the nine cells, not the three margins. Code-mixing is 16.1% low, 53.3% mid and 30.6% high; noise is 12.4% clean, 59.0% moderate and 28.6% noisy.

Domain

Logistics 260 calls, Marketplace 240, Travel 187, BFSI 122, Education 54, across 19 campaigns. No vertical exceeds 23% of the speech hours, and between 23% and 40% of each vertical's vocabulary appears in no other vertical, which is what makes the domain slices worth reading separately.

Speech versus silence

Four-fifths of this audio is not speech: ringing, hold, pauses and the agent's turn. This is the single biggest structural difference from clip-based read-speech sets, and it is deliberate. The honest size of this benchmark is 5.53 hours of speech, not 29.58 hours of audio, and that is the number quoted everywhere in this card.

Vocabulary

67,443 tokens and 4,086 distinct words. 40.7% of words appear exactly once and Yule's K is 89.8: conversational speech repeats a small set of function words, so the diagnostic numbers are the ones describing the tail. That shape carries a risk: a large, easy filler bucket can mask failures on rare domain words, which is why meaning-changing errors and their classes are reported next to the aggregate.

Speakers and conditions

146 calls are female speakers (16.9%), from human annotation, and 717 are male. The median call is 80 seconds (P10 38, P90 284) carrying 14 seconds of customer speech (4–51), spoken at a median 188 words per minute (118–276). Caller accent is not annotated.

Why we score meaning

Both of these turns contain exactly one wrong word.

Two reference turns with one wrong word each. 'तीन हजार लगभग' heard as 'मतलब तीन हजार लगभग' keeps the amount intact. 'छबीस' (26) heard as 'छत्तीस' (36) changes the number the customer said.

Reference	System output	Word errors	Judge's verdict	What happens next
`तीन हजार लगभग` (about three thousand)	`मतलब तीन हजार लगभग`	1 (one extra word)	harmless · filler	Nothing. The amount is right.
`छबीस` (twenty-six)	`छत्तीस` (thirty-six)	1 (one wrong word)	counts · number	The agent hears the wrong number.

Raw word error rate counts one error in each, so it scores them alike. A voice agent does not read the transcript, it acts on it. So the headline metric here is semantic WER: errors that could change the meaning of a turn in a way that affects the downstream task, divided by reference words. In everyday telesales speech, fillers and repetitions dominate the raw error count and rarely matter. For each error, the judge sees the reference turn, the system output and the error, answers would the agent act differently?, and tags a class (here filler and number).

Evaluation protocol

Scoring flow: reference normalisation, vendor normalisation and alignment to reference turns, deterministic error counting per turn, LLM judging of each error, semantic WER.

Three things about this protocol are load-bearing.

The same recordings and the same pipeline for every system. All 11 systems transcribed the same 863 calls, and one pinned model (Gemini 3.7 Flash, temperature 0) runs every LLM step with the same prompts. If a system returned nothing for a call, every reference word of that call counts as deleted.

The normalisers are ours, not yours. Submissions carry raw transcripts and both normalisation passes run on our side. If each vendor normalised its own output, the leaderboard would measure normalisation effort as much as recognition quality. Numeric values, negation, named entities and word order are never normalised: a wrong number is an error, a differently formatted number is not. हाँ, 15। becomes हां fifteen; रिक्वायरमेंट becomes requirement.

Each error is judged on its own. One error, one question (would the agent act differently?), answered inside the turn it occurred in, with no sight of the running total. Every error is then tagged with a class, and both the verdict and the class ship in results, so the judgement is auditable rather than asserted.

Aligned per turn, not per call

Vendors cut speech differently and sometimes drop a whole turn. Aligned over the whole call, a matcher can pair a word with its neighbour in the next turn. In one real call the reference says "fifty percent" and the system says "sixty percent"; whole-call alignment pairs हां (the previous turn) with "sixty" and marks "fifty" deleted, while per-turn alignment keeps it as one number error in the turn where it happened.

A vendor returning ten lines for a six-turn reference, aligned back to the reference turns.

On a 240-call subset, per-turn counting was 0.5 to 1.0 point stricter in token WER for five of the six systems tested, because words can no longer be matched across turns.

The metrics

Metric	Definition
Semantic WER (headline)	meaning-changing errors ÷ reference words. A voice agent does not read the transcript, it acts on it.
Raw WER	`(S+D+I)/N`, with the three components published separately so the filtering is auditable
Error class	each meaning-changing error tagged `number`, `negation`, `name`, `commitment` or `content`; the classes sum to semantic WER
Latency (FTR)	time from our end-of-turn finalize signal to the vendor's final reply for that turn, 80th percentile (P80) over all turns, on the locked 50-call set; TTFS adds a constant 480 ms (method)

Latency measurement

Each call is streamed in real time; at every end of a customer turn, decided by our production VAD (stop 0.5 s, an effective 0.48 s at 8 kHz), we send the system its finalize signal. FTR is the time from that signal to the system's reply for the turn, on a locked 50-call set at 5 concurrency, one system at a time, from one server in Delhi. TTFS = FTR + 480 ms: the time from actual speech end, as in pipecat-ai/stt-benchmark, which includes the VAD wait (Pipecat's benchmark defaults to a 0.2 s stop, so compare FTR, not TTFS, across the two). Full method, P50–P99 per system and the Soniox region comparison: docs/LATENCY.md.

Planned for the next release: confidence intervals and paired comparisons, character error rate, entity WER (errors divided by entity tokens), hallucination on silence, and a measured agreement between the judge and human adjudication.

Baseline results

Ordered by semantic WER, lowest first. SquadStack's in-house systems are marked ★ and ranked with the others: comparing a model tuned on this kind of audio with a general-purpose endpoint as peers is the usual way vendor benchmarks mislead.

System	Vendor	Category	Semantic WER	Raw WER	Sub	Del	Ins	FTR P80 (ms)
Soniox v5	Soniox	Commercial	8.77	27.29	12.66	5.45	9.17	89.4
Arth V2 ★	SquadStack	In-house	8.85	29.14	13.17	8.05	7.92	67.0
Arth V1 ★	SquadStack	In-house	9.17	29.83	13.42	8.96	7.45	67.0
Nova 3	Deepgram	Commercial	9.31	30.98	11.69	14.49	4.80	93.8
Saaras v4	Sarvam	Commercial	10.84	33.25	14.73	6.26	12.26	176.7
Cartesia ink-2	Cartesia	Commercial	11.58	36.27	14.69	3.84	17.73	183.6
Gemini 3.5 Transcribe	Google	Commercial	11.93	35.82	14.48	14.67	6.68	264.2
OpenAI gpt-live-transcribe	OpenAI	Commercial	12.30	38.14	18.16	13.52	6.46	830.8
Chirp 3	Google	Commercial	12.47	34.03	15.14	10.96	7.93	n/a
Flux	Deepgram	Commercial	14.33	37.76	18.28	12.91	6.56	124.4
Scribe v2 Realtime	ElevenLabs	Commercial	22.09	54.95	23.99	9.09	21.87	86.5

Pooled over about 67,000 reference words (67,388–67,393 per system); Sub, Del and Ins are percentages of reference words and add to raw WER. Latency: FTR P80 on the locked 50-call set (see above). Point estimates: systems within about a point of each other should not be ranked on this table alone.

What raw WER hides

For each system, raw WER split into the part that changes meaning and the part that does not, and into substitutions, deletions and insertions.

Between 30% and 40% of every system's raw word errors would change what an agent does; for the median system it is 32%. A 29.1% raw WER is an 8.8% rate of errors that matter, so a decision made on raw WER reads a gap roughly three times its real size.

The ordering mostly survives, and the exceptions are in the middle of the table. Chirp 3 is sixth on raw WER and ninth on semantic WER: 37% of its errors change meaning, against 30–32% for the four leading systems. Cartesia ink-2 moves up two places because about half of its raw errors are insertions that the judge mostly treats as harmless. OpenAI also moves up two, passing Flux and Chirp 3, which lose a larger share of their errors to meaning. The top five keep their places.

What the errors land on

Each meaning-changing error carries a class, and the classes sum to each system's semantic WER (percent of reference words):

System	Semantic WER	Number	Negation	Name	Commitment	Content
Soniox v5	8.77	1.07	0.43	0.77	0.30	6.18
Arth V2 ★	8.85	1.10	0.52	0.92	0.42	5.89
Arth V1 ★	9.17	1.11	0.52	0.99	0.41	6.14
Nova 3	9.31	1.24	0.55	1.11	0.48	5.92
Saaras v4	10.84	1.29	0.59	0.90	0.50	7.55
Cartesia ink-2	11.58	1.15	0.60	0.93	0.42	8.47
Gemini 3.5 Transcribe	11.93	1.43	0.79	0.95	0.65	8.08
OpenAI gpt-live-transcribe	12.30	1.59	0.68	1.02	0.46	8.55
Chirp 3	12.47	1.91	0.70	0.91	0.52	8.41
Flux	14.33	1.68	0.86	1.36	0.64	9.79
Scribe v2 Realtime	22.09	2.11	1.33	1.47	0.92	16.25

Two systems with the same semantic WER are not equally usable. Number and negation errors change the outcome of a call (a wrong amount is quoted, a decline is heard as a yes) while a content error usually costs a retry. Content dominates for every system; number errors run from 1.1% to 2.1% of reference words.

By slice

Heatmap of semantic WER for every system across noise, code-mixing, domain and gender slices.

Noise is the largest factor. Semantic WER rises 4–12 points from clean to noisy calls (Soniox 4.9 → 12.7, Cartesia 5.6 → 17.1).
Code-mixing costs less than noise. Going from low to high mixing adds 2–5 points.
Marketplace is the hardest vertical, Education the easiest. Soniox: 13.9% on Marketplace, 6.5% on Education (54 calls).
No consistent gender gap (female 8.1–12.7% vs male 8.8–14.7% for all but Scribe v2), on 146 female calls.

Uses

Direct use

Comparing speech-to-text systems for deployment in Indian consumer-sales voice agents. The claim this benchmark licenses you to make is narrow and specific:

On Hindi–English telesales calls over 8 kHz telephony, system A made meaning-changing transcription errors at rate X and returned a final transcript within Y ms of the end-of-turn signal for 80% of turns.

It also supports per-condition diagnosis: which systems degrade on heavy code-mixing or noise, which lose numbers, which flip negations.

Out-of-scope use

Training or fine-tuning any model. This is a licence violation, not a recommendation: see Licensing.
Speaker identification, speaker verification, voice cloning, or any attempt to re-identify a caller.
Claims about ASR quality in general, in other languages, in other domains, or on clean wide-band audio.
Claims about a vertical represented by few calls, such as Education (54 calls).
Procurement decisions on latency alone. Latency here is from one server location and a 50-call set; reproduce the measurement from your own network location before acting on it.

Licensing

Four separate statements, because they genuinely differ:

Component	Licence
Audio	Benchmark Evaluation Licence 1.0: evaluation only, no redistribution, no training, no speaker identification. Recordings remain the property of SquadStack and its clients.
Reference transcripts and annotations	CC BY-NC 4.0, attribution to SquadStack
Compilation, metadata and schema	CC BY 4.0
Evaluation code (harness and metrics, supplied on request)	Apache 2.0

Full text in LICENSE.md; the access terms you accept at the gate are in TERMS_OF_USE.md. Commercial evaluation is permitted; publishing results is encouraged; training is not permitted under any of the four.

Citation

@dataset{telesales_asr_benchmark_2026,
  title        = {SquadStack Conversational Streaming ASR Benchmark (8 kHz)},
  author       = {{SquadStack MLOps Team}},
  year         = {2026},
  version      = {1.0.0},
  publisher    = {SquadStack},
  url          = {https://huggingface.co/datasets/Squadstack/conversational-streaming-asr-benchmark},
  note         = {Evaluation-only benchmark of 863 real Indian telesales calls,
                  8 kHz narrowband, Hindi-English code-mixed}
}

SquadStack MLOps Team (2026). SquadStack Conversational Streaming ASR Benchmark (8 kHz) (Version 1.0.0) [Data set]. SquadStack.

Acknowledgements

Curation methodology builds on the code-mixing index of Gambäck & Das (2016), MATTR (Covington & McFall, 2010), DNSMOS (Reddy et al., Microsoft) and NISQA (Mittag et al., TU Berlin). The evaluation design draws on the Open ASR Leaderboard, and the semantic-WER framing and final-transcript latency clock of pipecat-ai/stt-benchmark.

Dataset card authors SquadStack MLOps Team

Downloads last month: 64