MioTTS のローカルマシンへのインストール

WindowsでCUDAのモードで実行した場合、reference_waveform が CPU 上（torch.FloatTensor）にあるのに対し、 コーデックモデルは CUDA 上（torch.cuda.FloatTensor）にあるため、デバイス不一致が発生しています。という状況になり、 参照音声を利用した場合に400エラーが発生する場合があります。
その場合は、コードを変更すると改善できる場合があります。

synthesizeメソッドの変更

    def synthesize(
        self,
        tokens: list[int] | torch.Tensor,
        reference_waveform: torch.Tensor | None = None,
        global_embedding: torch.Tensor | None = None,
        target_audio_length: int | None = None,
    ) -> torch.Tensor:
        if reference_waveform is None and global_embedding is None:
            raise ValueError("Either reference_waveform or global_embedding is required.")

        # Extract global embedding from reference waveform if provided
        if reference_waveform is not None:
            ref_features = self.codec.encode(reference_waveform, return_content=False, return_global=True)
            global_embedding = ref_features.global_embedding

        if isinstance(tokens, list):
            tokens = torch.tensor(tokens, dtype=torch.long, device=_codec_device(self.codec))
        elif isinstance(tokens, torch.Tensor) and tokens.dtype != torch.long:
            tokens = tokens.long()
        return self.codec.decode(
            global_embedding=global_embedding,
            content_token_indices=tokens,
            target_audio_length=target_audio_length,
        )

    def synthesize(
        self,
        tokens: list[int] | torch.Tensor,
        reference_waveform: torch.Tensor | None = None,
        global_embedding: torch.Tensor | None = None,
        target_audio_length: int | None = None,
    ) -> torch.Tensor:
        if reference_waveform is None and global_embedding is None:
            raise ValueError("Either reference_waveform or global_embedding is required.")

        device = _codec_device(self.codec)

        # Extract global embedding from reference waveform if provided
        if reference_waveform is not None:
            reference_waveform = reference_waveform.to(device=device, dtype=torch.float32)
            ref_features = self.codec.encode(reference_waveform, return_content=False, return_global=True)
            global_embedding = ref_features.global_embedding

        if isinstance(tokens, list):
            tokens = torch.tensor(tokens, dtype=torch.long, device=device)
        elif isinstance(tokens, torch.Tensor) and tokens.dtype != torch.long:
            tokens = tokens.long()
        return self.codec.decode(
            global_embedding=global_embedding,
            content_token_indices=tokens,
            target_audio_length=target_audio_length,
        )

synthesize_batchメソッドの変更

    def synthesize_batch(
        self,
        tokens_list: list[list[int]] | list[torch.Tensor],
        reference_waveform: torch.Tensor | None = None,
        global_embedding: torch.Tensor | None = None,
        target_audio_lengths: list[int] | None = None,
        padding_token_idx: int = 0,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if reference_waveform is None and global_embedding is None:
            raise ValueError("Either reference_waveform or global_embedding is required.")
        logger.debug("Synthesize batch: items=%d", len(tokens_list))

        device = _codec_device(self.codec)

        # Extract global embedding from reference waveform if provided
        if reference_waveform is not None:
            ref_features = self.codec.encode(reference_waveform, return_content=False, return_global=True)
            global_embedding = ref_features.global_embedding

    def synthesize_batch(
        self,
        tokens_list: list[list[int]] | list[torch.Tensor],
        reference_waveform: torch.Tensor | None = None,
        global_embedding: torch.Tensor | None = None,
        target_audio_lengths: list[int] | None = None,
        padding_token_idx: int = 0,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if reference_waveform is None and global_embedding is None:
            raise ValueError("Either reference_waveform or global_embedding is required.")
        logger.debug("Synthesize batch: items=%d", len(tokens_list))

        device = _codec_device(self.codec)

        # Extract global embedding from reference waveform if provided
        if reference_waveform is not None:
            reference_waveform = reference_waveform.to(device=device, dtype=torch.float32)
            ref_features = self.codec.encode(reference_waveform, return_content=False, return_global=True)
            global_embedding = ref_features.global_embedding

起動用のバッチファイルのデバイス指定を --device cuda に変更します。

call .venv\Scripts\activate.bat

start "OLLAMA SERVER" cmd /k ollama serve
timeout /t 3 /nobreak > nul

start "OLLAMA MODEL" cmd /k ^
ollama run hf.co/Aratako/MioTTS-GGUF:MioTTS-2.6B-BF16.gguf
timeout /t 5 /nobreak > nul

start "MIOTTS SERVER" cmd /k ^
uv run run_server.py --llm-base-url http://localhost:11434/v1 --device cuda
timeout /t 5 /nobreak > nul

start "MIOTTS WEBUI" cmd /k python run_gradio.py

この修正で改善できる場合があります。