deezer/spleeter · error · ValueError

T is too large considering STFT parameters and chunk duratoi

Error message

T is too large considering STFT parameters and chunk duratoin. Make sure spectrogram time dimension of chunks is larger than T (for instance reducing T or frame_step or increasing chunk duration).

What it means

The same compatibility check also validates the time dimension: `(chunk_duration * sample_rate - frame_length) / frame_step` must be >= T. If the number of STFT frames available in a chunk is smaller than T, a `ValueError` is raised because spectrogram chunks cannot supply T time steps for the model.

Source

Thrown at spleeter/dataset.py:307

        self._chunk_duration = chunk_duration
        self._audio_adapter = audio_adapter
        self._audio_params = audio_params
        self._audio_path = audio_path
        self._random_seed = random_seed

        self.check_parameters_compatibility()

    def check_parameters_compatibility(self):
        if self._frame_length / 2 + 1 < self._F:
            raise ValueError(
                "F is too large and must be set to at most frame_length/2+1. "
                "Decrease F or increase frame_length to fix."
            )

        if (
            self._chunk_duration * self._sample_rate - self._frame_length
        ) / self._frame_step < self._T:
            raise ValueError(
                "T is too large considering STFT parameters and chunk duratoin. "
                "Make sure spectrogram time dimension of chunks is larger than T "
                "(for instance reducing T or frame_step or increasing chunk duration)."
            )

    def expand_path(self, sample: Dict) -> Dict:
        """Expands audio paths for the given sample."""
        return dict(
            sample,
            **{
                f"{instrument}_path": tf.strings.join(
                    (self._audio_path, sample[f"{instrument}_path"]), SEPARATOR
                )
                for instrument in self._instruments
            },
        )

    def filter_error(self, sample: Dict) -> tf.Tensor:

View on GitHub (pinned to c8854001ac)

Solutions

1. Reduce T to fit: T <= (chunk_duration * sample_rate - frame_length) / frame_step
2. Increase chunk_duration so each chunk yields at least T frames
3. Reduce frame_step (finer hop) to increase frames per chunk

Example fix

// before
SpleeterDataset(chunk_duration=1.0, sample_rate=44100, frame_length=2048, frame_step=1024, T=64, ...)  # ~41 frames < T
// after
SpleeterDataset(chunk_duration=2.0, sample_rate=44100, frame_length=2048, frame_step=1024, T=64, ...)  # ~84 frames >= T

def validate_stft_t(chunk_duration: float, sample_rate: int, frame_length: int, frame_step: int, T: int):
    frames = (chunk_duration * sample_rate - frame_length) / frame_step
    if frames < T:
        raise ValueError(f'T must be <= {int(frames)} (got T={T}); increase chunk_duration or reduce T/frame_step')
validate_stft_t(chunk_duration=2.0, sample_rate=44100, frame_length=2048, frame_step=1024, T=64)

try:
    dataset = SpleeterDataset(chunk_duration=chunk_duration, T=T, frame_step=frame_step, ...)
except ValueError as e:
    if 'T is too large' in str(e):
        T = int((chunk_duration * sample_rate - frame_length) / frame_step)
        dataset = SpleeterDataset(chunk_duration=chunk_duration, T=T, frame_step=frame_step, ...)
    else:
        raise

When it happens

Trigger: Constructing `SpleeterDataset(...)` with T larger than the frame count implied by chunk_duration, sample_rate, frame_length, and frame_step — e.g. very small chunk_duration or very large frame_step relative to T.

Common situations: Shortening chunk_duration to save memory without reducing T, changing sample_rate or frame_step in a config inherited from another model, copying hyperparameters between datasets with different chunk lengths.

Related errors

• F is too large and must be set to at most frame_length/2+1.
• {adapter_class_name} is not a valid AudioAdapter class
• No model function {model_type} found
• Unkwnown loss type: {loss_type}
• Invalid mask_extension parameter {extension}

AI-assisted analysis of deezer/spleeter@c8854001ac (2026-08-28). Data as JSON: /api/errors/15be750f6cc49507. Report an issue: GitHub.