Speech "cuts off" at the end of each sentence

[Thelukepet]

Hey, I tried reducing the speed for the .generate_speech and increasing the duration of the '.encode_prompt` but it did not fix the issue when using english samples + ref_text.

In the end I fixed it by changing the generate function in the modeling_utils.py, specifically by adding padding to the Vocoder to give it enough data to finish the previous sound. Anything else I should have looked at?

def generate(
    prompt_tokens,
    prompt_features_lens,
    prompt_features,
    prompt_rms,
    text,
    model,
    vocoder,
    tokenizer,
    num_step=4,
    guidance_scale=3.0,
    speed=1.0,
    t_shift=0.5,
    target_rms=0.1,
):
    tokens = tokenizer.texts_to_token_ids([text])
    device = next(model.parameters()).device  # Auto-detect device

    speed = speed * 1.3

    with torch.inference_mode():
        (pred_features, _, _, _) = model.sample(
            tokens=tokens,
            prompt_tokens=prompt_tokens,
            prompt_features=prompt_features,
            prompt_features_lens=prompt_features_lens,
            speed=speed,
            t_shift=t_shift,
            duration="predict",
            num_step=num_step,
            guidance_scale=guidance_scale,
        )

    # Convert to waveform
    pred_features = pred_features.permute(0, 2, 1) / 0.1

    # FIX: Padding for the Vocoder
    # We take the last frame and repeat it 15 times (approx 150ms buffer)
    # This gives Vocos enough data to finish the previous sound without cutting it.
    last_frame = pred_features[:, :, -1:]
    padding_frames = last_frame.repeat(1, 1, 15)
    pred_features = torch.cat([pred_features, padding_frames], dim=2)

    wav = vocoder.decode(pred_features).squeeze(1).clamp(-1, 1)

    # Volume matching
    if prompt_rms < target_rms:
        wav = wav * (prompt_rms / target_rms)

    return wav