Fix Base voice clone streaming quality and stop-token crash

[linyueqian]

Summary

• Fix voice clone (Base ICL) audio distortion in async_chunk streaming mode
• Fix CUDA assert crash when using async_chunk=false with voice clone
• Align context trimming with official HF qwen_tts implementation

Changes

1. Streaming quality (stage_input_processors/qwen3_tts.py)

Prepend ref_code as decoder context on every chunk, not just the first. The HF reference decodes ref_code + all_codes in one pass; without ref_code on later chunks the vocoder loses speaker identity and produces distorted audio. Changed .pop() to .get() to retain ref_code across chunks.

2. Stop-token crash (stage_input_processors/qwen3_tts.py)

Filter frames with codec values >= codebook_size (2048). The talker's stop_token_id (2150) exceeds the codebook and causes CUDA assert in the decoder's embedding lookup when using the non-streaming path.

3. Proportional context trim (qwen3_tts_code2wav.py)

Use ctx_frames / actual_frames * wav_len ratio (matching HF's qwen_tts implementation) instead of fixed ctx_frames * upsample to handle decoder output length variations.

Test plan

• Base (voice clone) with async_chunk=true produces clean audio matching HF quality
• Whisper transcription confirms correct text content
• CustomVoice and VoiceDesign still work (no regression)
• Base with async_chunk=false no longer crashes

Fixes #1944

CC @Sy0307

[linyueqian]

Test Results

Text: "Good one. Okay, fine, I'm just gonna leave this sock monkey here. Goodbye."
Model: Qwen/Qwen3-TTS-12Hz-1.7B-Base (voice clone)
Ref audio: clone_2.wav from official Qwen3-TTS repo

A/B Comparison

Version	Duration	Whisper Transcription	Audio Quality
HF qwen_tts (reference)	4.80s	Good. Okay, fine. I'm just gonna leave this sock monkey here. Goodbye.	Clean, no distortion
vLLM before fix	5.04s	Good one. Okay, fine. I'm just gonna leave this sock monkey here. Goodbye.	Distorted after first few seconds
vLLM after fix	5.04s	Good one. Okay, fine. I'm just gonna leave this sock monkey here. Goodbye.	Clean, matches HF quality

Root Cause

Before fix: Only the first streaming chunk received ref_code decoder context (101 frames). Subsequent chunks lost speaker identity → distorted audio.

After fix: Every chunk receives ref_code as left context, maintaining speaker identity throughout the stream — matching HF's one-pass decode behavior.

Environment

• A100-SXM4-80GB, CUDA 12.4, vLLM 0.17.0
• gpu_memory_utilization=0.08 (shared GPU)
  hf.wav
  vllm_before_fix.wav
  vllm_after_fix.wav

[hsliuustc0106]

can we add this to the acc test for qwen-tts model

[linyueqian]

can we add this to the acc test for qwen-tts model

voice clone is hard to test given the timbre is easy to be identified from human, but ASR cannot tell any difference. @yenuo26 any good suggestion? i notice you add gender test which seems great.

[linyueqian]

wespeaker could be one option

[hsliuustc0106]

lgtm

[hsliuustc0106]

wespeaker could be one option

can we have a minimal assertion to avoid "Distorted after first few seconds"?

[linyueqian]

Audio Quality Assertion

Added a minimal HNR (Harmonic-to-Noise Ratio) check that reliably detects the voice clone distortion:

def assert_voice_clone_quality(wav_path, min_hnr_db=1.2):
    """Distorted voice clone: HNR < 1.0 dB. Clean: HNR > 1.2 dB."""
    data, sr = sf.read(wav_path)
    frame_len = int(0.03 * sr)
    hop = frame_len // 2
    hnr_values = []
    for start in range(0, len(data) - frame_len, hop):
        frame = data[start:start + frame_len]
        if np.max(np.abs(frame)) < 0.01:
            continue
        ac = np.correlate(frame, frame, mode='full')[len(frame)-1:]
        ac = ac / (ac[0] + 1e-10)
        min_lag, max_lag = int(sr / 400), min(int(sr / 80), len(ac))
        if min_lag >= max_lag:
            continue
        peak = np.max(ac[min_lag:max_lag])
        if 0 < peak < 1:
            hnr_values.append(10 * np.log10(peak / (1 - peak + 1e-10)))
    return np.mean(hnr_values) if hnr_values else 0

Results

Sample	HNR	Status
Before fix (distorted)	0.88 dB	FAIL
After fix	1.62 dB	PASS
HF reference	2.17 dB	PASS
After fix sample 1	4.38 dB	PASS
After fix sample 2	4.88 dB	PASS
After fix sample 3	1.65 dB	PASS
After fix sample 4	1.45 dB	PASS
After fix sample 5	2.85 dB	PASS

Threshold of 1.2 dB cleanly separates distorted from clean voice clone audio across all test cases.

[univa-HARRY]

I think it would resolve my issue #1707

[yenuo26]

can we add this to the acc test for qwen-tts model

voice clone is hard to test given the timbre is easy to be identified from human, but ASR cannot tell any difference. @yenuo26 any good suggestion? i notice you add gender test which seems great.

I have added audio content generation and voice gender validation in the use cases. Currently, I'm using F0 for simple calculation, but I feel this approach is less stable. I'm currently considering switching to a small gender recognition model.
#1911

[Sy0307]

LGTM. Tests go well. I will save ref_code in metadata instead of prompt_token_ids in a new PR.