srfd-voxcpm2-demo

SRFD-VoxCPM2 Demo

Private demo bundle for SR-FD VoxCPM2 few-step TTS experiments.

This repository packages:

A static demo website under site/
Aligned Seed-TTS English audio comparisons under site/audio/
Seed-TTS English benchmark results under data/results.json
The compact 3-target SR-FD LoRA adapter under model/ after validation
Reproduction configs under configs/
Utility scripts under scripts/

The model is a LoRA adapter for openbmb/VoxCPM2, not a standalone full checkpoint. Use it together with the VoxCPM2 base model.

Verified Results

The compact 3-target SR-FD adapter is packaged in model/. It obtains 167 word errors on upstream Seed-TTS English. The table keeps the paper-facing denominator 11805 for consistency with the main report.

System	Steps	FT	SR-FD	Upstream WER	UTMOS / DNSMOS OVRL / P808
Base VoxCPM2	4	No	No	263/11805 = 2.2279%	3.2974 / 2.8950 / 3.5296
Base VoxCPM2	10	No	No	205/11805 = 1.7366%	3.8072 / 3.0866 / 3.6689
Matched FT	4	Yes	No	174/11805 = 1.4740%	3.7615 / 3.0729 / 3.6522
FT + SR-FD	4	Yes	Yes	167/11805 = 1.4147%	3.7637 / 3.0711 / 3.6507
ARCHI-TTS reported	4 / few-step	-	-	1.47%	-

Compact 3-Target SR-FD

The compact target set keeps only content-centered SR-FD targets:

asr_true4_good_whisper: Whisper content statistics from ASR-reranked good true-4 generations.
teacher_t10_ctc_content: CTC posterior statistics from ten-step teacher generations.
real_ctc_content: CTC posterior statistics from real LibriTTS voice-cloning speech.

This design keeps the paper story focused on few-step intelligibility while removing auxiliary quality and WavLM targets from the main model.

For the compact run, SR-FD is applied after length gating without row-level sample-weight gating. This avoids depending on internal manifest weight fields and makes the compact loss easier to reproduce.

The selected compact checkpoint is fourstep_fdcompact3nogate_bal_fd2e4_lr3e8_s42/step_0001600. Among the promoted compact variants, it is the only one that reaches 167 upstream word errors. The other promoted variants score 173, 176, and 182 errors.

Local Demo

Open the static site directly:

python3 -m http.server 8080 --directory site

Then visit http://localhost:8080.

The Audio section compares the same Seed-TTS prompts across:

Base VoxCPM2 4-step
Base VoxCPM2 10-step
Matched 4-step fine-tuning without SR-FD
FT + SR-FD
Three compact leave-one-out ablations

Each example includes prompt reference audio, target reference audio, and the reference transcript used for WER. Each audio card for every model includes the generated waveform, reference transcript, ASR transcript, per-utterance WER, and full-set WER. Zero-WER FT + SR-FD examples are therefore inspectable as exact reference/ASR transcript matches, not just aggregate scores.

The Audio section also includes a Negative Cases tab. Those examples are selected from full Seed-TTS evaluation rows where FT + SR-FD still makes a word-level ASR mistake or regresses against matched fine-tuning, so the demo shows both the frontier behavior and the remaining failure modes.

Adapter Layout

model/ contains:

model/
  lora_config.json
  lora_weights.safetensors
  training_state.json
  selected_checkpoint.json

Only inference-relevant adapter files are included. Optimizer, scheduler, and SR-FD queue state are excluded from the demo bundle.

License

This demo follows the base model’s Apache-2.0 license terms. See openbmb/VoxCPM2 for the original model card and usage restrictions.