Rcarvalo/vibevoice-lora-siwis-samples

VibeVoice-Realtime-0.5B → SIWIS French voice transfer (LoRA)

End-to-end artefacts for shifting microsoft/VibeVoice-Realtime-0.5B from its built-in fr-Spk0_man.pt (MS Quebec) French voice to a metropolitan / SIWIS style voice, via a tiny LoRA on tts_language_model.

Contains:

• Audio samples (8 test cases × {baseline, LoRA}) — what to listen to
• lora_best.safetensors — the trained LoRA adapter (6 MB, rank 16)
• aligned_encoder.safetensors — the aligned 0.5B ↔ 1.5B encoder used to pre-tokenize SIWIS (1.37 GB)
• scripts/ — the exact training + inference + diagnostic code we ran

TL;DR : LoRA on VibeVoice-Realtime only works if training mirrors the inference pipeline. Concretely, you must load an MS voice prompt as past_key_values prefix at training time. Training without a prefix produces a LoRA whose weight deltas look fine (~1 % relative) but have zero audible effect because the training regime is out-of-distribution vs. inference.

---

1. What we were trying to do

VibeVoice-Realtime-0.5B is a zero-shot voice-cloning TTS : you pass it a voice_prompt.pt (a precomputed KV cache of a reference audio) plus text, and it generates speech in that voice. It ships with 25 voice prompts across 13 languages. The built-in French prompt fr-Spk0_man.pt clearly sounds Québécois to a European French ear, and we wanted a metropolitan French voice closer to SIWIS.

Two obvious paths :

1. Build a new voice prompt by passing a SIWIS reference audio through the encoder + language_model + tts_language_model prefill flow → save the resulting KV cache as a .pt.
2. Fine-tune the model (LoRA) so that when it sees an existing MS voice prompt, it generates SIWIS-style output.

Path 1 is the "proper" way but internally VibeVoice's prefill flow for tts_language_model is non-trivial to reproduce (see microsoft/VibeVoice#117 and the multiple create_voice_prompt_v{2..11}.py attempts in the Rcarvalo repo) : our caches were statistically close to MS Quebec but diverged layer-by-layer by 3–14 %, giving broken audio.

Path 2 → this repo.

---

2. First attempt : why it failed (the lesson)

Setup

Standard LoRA on attention q/k/v/o of tts_language_model, rank 8, LR 3e-5, 10 epochs, 910 optimizer steps. Training forward was exactly the same as the pre-existing training/vibevoice/train.py in the repo :

lm_out  = model.model.language_model(inputs_embeds=text_embeds, use_cache=False)
tts_out = model.model.tts_language_model(inputs_embeds=seq_embeds, use_cache=False)
loss    = ddpm_v_prediction_loss(..., tts_hidden, target_latents, ...)

Loss dropped 0.79 → 0.49 (val 0.4921), which looked healthy.

Inference gave zero audible change

We generated the 5 SIWIS test phrases. The LoRA-merged model produced audio indistinguishable from the base model — still Quebec.

Diagnostic

Ran scripts/diag_lora_merge.py, which snapshots 3 target weights before merging and compares after :

Adapter state: 280 tensors (140 lora_A + 140 lora_B)
lora_B max_abs = 0.0137      (non-zero ⇒ training did update)
Merge: 140/140 layers merged, 0 missing

tts_language_model.layers.0.self_attn.q_proj.weight
    delta norm = 0.148  (rel 0.8 %)
tts_language_model.layers.10.self_attn.q_proj.weight
    delta norm = 0.109  (rel 0.6 %)
tts_language_model.layers.19.mlp.down_proj.weight
    delta norm = 0.451  (rel 1.5 %)

So :

• LoRA weights did train (non-zero B)
• Merge into the base model did happen (140/140)
• Weight deltas are real but small (~1 % relative)

Something real changed in the weights yet no audible effect. Why ?

Root cause : training-inference regime mismatch

At inference, VibeVoice.generate() does :

all_prefilled_outputs = copy.deepcopy(voice_prompt)   # ~323 tokens of frozen KV cache
outputs = model.generate(**inputs, all_prefilled_outputs=...)

The new tokens are prefilled on top of the voice prompt's past_key_values. The attention patterns for new tokens therefore have ~323 "context" K/V positions that were computed by the base model at MS training time — completely out-of-distribution vs. what our LoRA saw during training (use_cache=False, position 0, no prefix).

Our LoRA's updated q_proj projects new queries that attend over :

• The MS-Quebec cached K (untouched by LoRA — cached values are frozen .pt data)
• Newly computed K for the new tokens (LoRA-modified)

The LoRA was optimized to produce good latents on sequences starting from scratch, not on sequences anchored by a 323-token prefix. At inference, it operates OOD and its effect washes out.

---

3. What made the difference : prefix-conditioned training

One-line fix : at training time, load an FR voice prompt as past_key_values for both language_model and tts_language_model, and continue position_ids past the prefill length. Now training attention patterns look exactly like inference attention patterns.

Concretely, the v2 training forward

See scripts/train_streaming_lora.py :: training_forward_with_prefix for the canonical implementation. The only changes vs. v1 are marked ⭐ :

# ⭐ Clone the FR voice prompt's KV cache (one per model sub-module)
lm_past  = clone_cache(voice_prompt["lm"]["past_key_values"])
tts_past = clone_cache(voice_prompt["tts_lm"]["past_key_values"])
lm_prefill_len  = voice_prompt["lm"]["prefill_len"]    # e.g. 118 tokens
tts_prefill_len = voice_prompt["tts_lm"]["prefill_len"]  # e.g. 323 tokens

# --- language_model on new text, with FR prefix cache ---
text_embeds = model.model.get_input_embeddings()(text_ids)
# ⭐ Position IDs continue past the prefill
lm_position_ids = torch.arange(lm_prefill_len, lm_prefill_len + T_text, ...)
# ⭐ Attention mask covers prefill + new text
lm_attn_mask = torch.cat([torch.ones(B, lm_prefill_len), text_mask], dim=1)
lm_out = model.model.language_model(
    inputs_embeds=text_embeds,
    attention_mask=lm_attn_mask,
    position_ids=lm_position_ids,
    past_key_values=lm_past,   # ⭐
    use_cache=True,             # ⭐
)
new_lm_hidden = lm_out.last_hidden_state

# --- tts_language_model on [new_text_hidden | speech_embs], with tts prefix cache ---
speech_embeds = model.model.acoustic_connector(speech_norm)
seq_embeds = torch.cat([new_lm_hidden, speech_embeds], dim=1)
seq_embeds = seq_embeds + model.model.tts_input_types(type_ids)
tts_position_ids = torch.arange(tts_prefill_len, tts_prefill_len + T_total, ...)
tts_attn_mask = torch.cat([torch.ones(B, tts_prefill_len), combined_mask], dim=1)
tts_out = model.model.tts_language_model(
    inputs_embeds=seq_embeds,
    attention_mask=tts_attn_mask,
    position_ids=tts_position_ids,
    past_key_values=tts_past,   # ⭐ LoRA now learns to attend FR prefix context
    use_cache=False,             # don't accumulate
)

Only the last call has grad enabled (the LoRA lives inside tts_language_model). Everything upstream (get_input_embeddings, language_model, acoustic_connector, tts_input_types) runs under no_grad.

Loss is the same DDPM v-prediction on the speech-position hidden states as in train_base.py.

Hyperparameters (final run)

Knob	Value	Notes
LoRA targets	q_proj, k_proj, v_proj, o_proj of tts_language_model	attention only, no MLP (to limit drift)
Rank / alpha	16 / 32 (scaling = 2.0)
Trainable params	1.80 M / 675.73 M (0.267 %)
Optimizer	AdamW, betas (0.9, 0.95), weight decay 0.01
LR	5e-5, 30-step linear warmup, cosine to 10 %
Epochs	3 (540 opt steps, grad accum 8)
Prefix sources	fr-Spk0_man.pt, fr-Spk1_woman.pt (FR only)	random per step
Dataset	SIWIS French, 1444 train / 159 val	pre-tokenized with aligned_encoder.safetensors
GPU	1 × RTX A5500 (24 GB)	~6 min wall-clock

Loss trajectory (3 epochs)

Step	Train (batch)	Val
50	0.70	0.6704 (first best)
100	0.62	0.5871
150	0.58	0.5572
200	0.54	0.5425
300	0.53	0.5319
500	0.52	0.5187 (best)

Δ val 0.67 → 0.52 (-0.15) — descent is continuous and not plateauing, which is a clear sign that the gradients are finally hitting the right components (vs. the v1 run where val loss barely moved once fitted to the no-prefix regime).

---

4. Results

Quantitative bench (500 SIWIS French phrases, full table in BENCHMARK.md)

Metric	Baseline	LoRA SIWIS	Δ
WER mean	16.31 %	15.19 %	−1.12 pp ✅
RTF mean	0.4262	0.4302	+0.004 (≈0)
NISQA-MOS mean	2.848	3.770	+0.922 ✅✅

Headline: no degradation on WER or RTF, +0.92 on NISQA-MOS. See BENCHMARK.md for the full breakdown and raw per-phrase CSVs under benchmark/.

Audible test (listen to the wav pairs)

• FR (4 cases) : voice clearly shifts from Quebec to SIWIS timbre ✅
• EN / JP / IT (4 cases) : speech remains intelligible but the LoRA imposes the SIWIS timbre on top of the content — the multilingual speaker diversity is lost. Acceptable for a FR-only deployment, not for a multilingual app. Mitigations :
  • Reduce rank to 8 (less capacity ⇒ less drift on non-FR contexts)
  • Reduce merge scaling at inference from 2.0 to ~1.0
  • Add non-FR samples in training for regularization (requires EN/etc. target data)

Latency / baseline guarantees

LoRA is additive. With scaling 0.0 the model is exactly the base model bit-for-bit. With scaling 2.0 we get the SIWIS transfer. Merged weights on disk are standard .safetensors — no extra inference latency, no architectural change. Rolling back = reloading the base from_pretrained.

---

5. Repro

# 1) Install
git clone https://github.com/microsoft/VibeVoice.git && cd VibeVoice
pip install -e .[streamingtts]
# → gets transformers==4.51.3, vibevoice editable

# 2) Fetch SIWIS and pre-tokenize with the aligned encoder
python scripts/tokenize_data.py \
    --audio-dir /path/to/SiwisFrenchSpeechSynthesisDatabase/wavs \
    --transcripts /path/to/siwis_transcripts.jsonl \
    --encoder-path aligned_encoder.safetensors \
    --output-dir data/tokenized_siwis

# Split val off based on val_stems.txt (any ~10 % subset works)
python -c "
import shutil; from pathlib import Path
val_stems = set(Path('val_stems.txt').read_text().split())
src = Path('data/tokenized_siwis'); dst = Path('data/tokenized_siwis_val'); dst.mkdir(exist_ok=True)
for p in src.glob('*.pt'):
    if p.stem in val_stems: shutil.move(str(p), str(dst / p.name))
"

# 3) Train — 3 epochs, ~6 min on a single A5500
python scripts/train_streaming_lora.py \
    --data-dir data/tokenized_siwis \
    --val-dir data/tokenized_siwis_val \
    --output-dir outputs/lora_siwis \
    --epochs 3 --batch-size 1 --grad-accum 8 \
    --lr 5e-5 --warmup-steps 30 \
    --lora-rank 16 --lora-alpha 32

# 4) Smoke test + generate samples
python scripts/gen_samples_lora.py \
    --lora outputs/lora_siwis/lora_best.safetensors \
    --out-dir samples/ --lora-rank 16 --lora-alpha 32

# 5) Anti-degradation check (fr/en/jp/it before vs after)
python scripts/test_anti_degradation.py \
    --lora outputs/lora_siwis/lora_best.safetensors \
    --out-dir samples_antidegrade/ --lora-rank 16 --lora-alpha 32

The scripts are standalone and import only vibevoice, transformers, torch, safetensors, torchaudio. No custom training framework.

---

6. File inventory

README.md                          # this file

baseline_<label>.wav               # 8 × base VibeVoice outputs
lora_<label>.wav                   # 8 × LoRA-merged outputs
                                   # labels: fr_short_male, fr_medium_male,
                                   #         fr_practical, fr_female,
                                   #         en_short, en_medium, jp_short, it_short

lora_best.safetensors              # 6 MB — rank-16 LoRA for tts_language_model
                                   # (q/k/v/o on 20 layers = 80 LoRALinear)
aligned_encoder.safetensors        # 1.37 GB — aligned 1.5B encoder for 0.5B decoder
                                   # (needed to pre-tokenize SIWIS; see alignment/align_encoder_v3.py
                                   # in the Microsoft VibeVoice repo)

scripts/
    train_streaming_lora.py        # ⭐ v2 prefix-conditioned training (the fix)
    train_base.py                  # original no-prefix train.py (v1, failed — kept for reference)
    tokenize_data.py               # SIWIS wav → .pt via aligned encoder
    gen_samples_lora.py            # inference helper that merges LoRA at runtime
    test_anti_degradation.py       # generates 8 (baseline, lora) pairs across languages
    diag_lora_merge.py             # sanity check: does merging actually change weights ?

---

7. Credits & data

• Base model : microsoft/VibeVoice-Realtime-0.5B
• SIWIS : SIWIS French Speech Synthesis Database, Univ. of Edinburgh
• Aligned encoder training reference : microsoft/VibeVoice/alignment