Enabling stereo output of acc (Code Suggestion)

[Tom-M-Git]

The WebUI SVC pipeline currently collapses audio to mono before preprocessing and again during mixing, even though the separator supports stereo.

After modifying the code to preserve stereo:

>>> import soundfile as sf
>>> x, sr = sf.read("acc.wav")
>>> print(x.shape)
(663617, 2)

This shows the separator can output stereo when given stereo input.

In the current implementation:

• _trim_and_save_audio() uses librosa.load(..., mono=True)
• final mix also uses mono=True

This collapses stereo input before it reaches the separator.

Effect:

• Accompaniment loses stereo width
• Final output becomes fully mono

Preserving stereo significantly improves listening quality (wide instrumental + centered vocal).

Suggested changes:

1. Preserve stereo in input:
   librosa.load(..., mono=False)

2. Handle shape correctly:

   • (channels, samples) → transpose before saving

3. Avoid forcing mono in final mix:
   librosa.load(..., mono=False)

4. Duplicate mono vocal to stereo before mixing

My Code:

def _trim_and_save_audio(src_audio_path: str, dst_wav_path: Path, max_sec: int, sr: int = SAMPLE_RATE) -> None:
    audio_data, _ = librosa.load(src_audio_path, sr=sr, mono=False)

    max_samples = max_sec * sr

    if audio_data.ndim == 1:
        audio_data = audio_data[:max_samples]
    else:
        audio_data = audio_data[:, :max_samples]
        audio_data = audio_data.T

    dst_wav_path.parent.mkdir(parents=True, exist_ok=True)
    sf.write(dst_wav_path, audio_data, sr)

					mix_sr = self.svc_config.audio.sample_rate
					vocal, _ = librosa.load(str(generated), sr=mix_sr, mono=False)
					acc, _ = librosa.load(str(acc_path), sr=mix_sr, mono=False)
					
					if vocal.ndim == 1:
					    vocal = np.stack([vocal, vocal], axis=0)
					if acc.ndim == 1:
					    acc = np.stack([acc, acc], axis=0)

					if acc_shift != 0:
					    acc = np.stack([
					        librosa.effects.pitch_shift(acc[ch], sr=mix_sr, n_steps=acc_shift)
					        for ch in range(acc.shape[0])
					    ], axis=0)
					    print(f"Applied pitch shift of {acc_shift} semitones to accompaniment to match vocal shift of {vocal_shift} semitones.")

					mix_len = min(vocal.shape[-1], acc.shape[-1])
					if mix_len > 0:
					    mixed = vocal[:, :mix_len] + acc[:, :mix_len]
					    peak = float(np.max(np.abs(mixed))) if mixed.size > 0 else 1.0
					    if peak > 1.0:
					        mixed = mixed / peak
					    mixed_path = save_dir / "generated_mixed.wav"
					    sf.write(str(mixed_path), mixed.T, mix_sr)
					    generated = mixed_path

---

Caveat

It slightly changes the separation result.
For example, if there is an echo of vocal panning left-to-right, in my case, it wasn't treated as vocal and remained in acc.

Maybe there is a way to keep the prompt mono anyway then sparate vocal more reliably, while turning target acc stereo?

[jlqian98]

Thanks for your great suggestion. We will consider adding a stereo output option in future updates.