RVC-Boss/GPT-SoVITS · error · ValueError

搜索音频长度 ({search_len}) 必须大于等于参考音频长度 ({ref_len})

Error message

搜索音频长度 ({search_len}) 必须大于等于参考音频长度 ({ref_len})

What it means

ValueError from a torch.jit.script-ed helper find_best_audio_offset_fast used in the v2Pro streaming path: it computes the best alignment offset of a reference clip inside a longer search window via conv1d cross-correlation, which is only defined when search_audio is at least as long as reference_audio. The guard runs inside the scripted function, so shape violations fail deterministically instead of producing a confusing conv1d kernel-size error.

Source

Thrown at GPT_SoVITS/stream_v2pro.py:193

        return refer, sv_emb

    def extract_latent(self, ssl_content):
        codes = self.vq_model.extract_latent(ssl_content)
        return codes[0]

    def forward(self, pred_semantic, text_seq, refer, sv_emb=None):
        return self.vq_model(
            pred_semantic, text_seq, refer, speed=1.0, sv_emb=sv_emb
        )[0, 0]


@torch.jit.script
def find_best_audio_offset_fast(reference_audio: Tensor, search_audio: Tensor):
    ref_len = len(reference_audio)
    search_len = len(search_audio)

    if search_len < ref_len:
        raise ValueError(
            f"搜索音频长度 ({search_len}) 必须大于等于参考音频长度 ({ref_len})"
        )

    # 使用F.conv1d计算原始互相关
    reference_flipped = reference_audio.unsqueeze(0).unsqueeze(0)
    search_padded = search_audio.unsqueeze(0).unsqueeze(0)

    # 计算点积
    dot_products = F.conv1d(search_padded, reference_flipped).squeeze()

    if len(dot_products.shape) == 0:
        dot_products = dot_products.unsqueeze(0)

    # 计算参考音频的平方和
    ref_squared_sum = torch.sum(reference_audio**2)

    # 计算搜索音频每个位置的平方和（滑动窗口）
    search_squared = search_audio**2

View on GitHub (pinned to d523079fc0)

Solutions

1. Ensure every search window passed in is >= the reference/overlap length — pad the final chunk with zeros to at least len(reference_audio) before the call.
2. Skip cross-correlation for tail chunks shorter than the overlap and use them as-is (or truncate the reference to the chunk length).
3. Double-check argument order and that both tensors are at the same sample rate/dtype.
4. If you control overlap_len, reduce it so it never exceeds your minimum chunk size.

Example fix

# before
best = find_best_audio_offset_fast(ref_chunk, search)  # ValueError when search shorter

# after
if len(search) < len(ref_chunk):
    pad = torch.zeros(len(ref_chunk) - len(search), dtype=search.dtype)
    search = torch.cat([search, pad])
best = find_best_audio_offset_fast(ref_chunk, search)

assert search_audio.shape[-1] >= reference_audio.shape[-1], (
    f"search {search_audio.shape[-1]} < reference {reference_audio.shape[-1]}"
)

def valid_sola_pair(ref: torch.Tensor, search: torch.Tensor) -> bool:
    return search.numel() >= ref.numel() and ref.dtype == search.dtype

When it happens

Trigger: Calling find_best_audio_offset_fast(reference_audio, search_audio) (directly or via the v2Pro streaming decode that hunts overlap positions) with len(search_audio) < len(reference_audio) — e.g. the SOLA overlap window handed in is shorter than the overlap template taken from the previous chunk.

Common situations: Custom chunk sizes/overlap_len that make the tail chunk smaller than the overlap template; final chunk of a stream (is_final) shorter than the reference overlap; callers swapping argument order; sample-rate mismatch making lengths inconsistent.

Related errors

• prompt_text cannot be empty when using SoVITS_V3
• SoVits V3/4模型不支持流式推理模式

AI-assisted analysis of RVC-Boss/GPT-SoVITS@d523079fc0 (2026-08-15). Data as JSON: /api/errors/3261d102171172fe. Report an issue: GitHub.