ERROR during transcription: cannot reshape tensor of 0 elements into shape [-1, 0] because the unspecified dimension size -1 can be any value and is ambiguous

[olegkirianov]

Hello, getting this error for some audio files. Most files are processed fine, but some cause this issue:

ERROR during transcription: cannot reshape tensor of 0 elements into shape [-1, 0] because the unspecified dimension size -1 can be any value and is ambiguous

File "/home/kiryanovoa/GigaAM/gigaam/preprocess.py", line 86, in forward
return self.featurizer(input_signal), self.out_len(length)
File "/home/kiryanovoa/GigaAM/gigaam/model.py", line 33, in forward
features, feature_lengths = self.preprocessor(features, feature_lengths)
File "/home/kiryanovoa/GigaAM/gigaam/model.py", line 159, in transcribe_longform
encoded, encoded_len = self.forward(wav, length)
File "/home/kiryanovoa/recognizer/transcribe.py", line 56, in transcribe_channels
result2 = model.transcribe_longform(channel_files[1])
File "/home/kiryanovoa/recognizer/transcribe.py", line 79, in transcribe_audio
transcription = transcribe_channels(channel_files)
File "/home/kiryanovoa/recognizer/test.py", line 10, in
transcription = transcribe_audio(file_location)
RuntimeError: cannot reshape tensor of 0 elements into shape [-1, 0] because the unspecified dimension size -1 can be any value and is ambiguous

Environment:
Python 3.10.12
NumPy: 2.2.6
Torch: 2.5.1+cu124
Torchaudio: 2.5.1+cu124
Onnx: 1.17.0

[KonstantinShabaev]

добавь строчку в model.py после length = torch.full([1], wav.shape[-1], device=self._device)
if lenth>0:
и должно сработать, это из-за того что vad_utils возвращает тензор нулевой длины
в итоге цикл в функции должен таким получиться:

       for segment, segment_boundaries in zip(segments, boundaries):
            wav = segment.to(self._device).unsqueeze(0).to(self._dtype)
            length = torch.full([1], wav.shape[-1], device=self._device)
            if length>0:
                encoded, encoded_len = self.forward(wav, length)
                result = self.decoding.decode(self.head, encoded, encoded_len)[0]
                transcribed_segments.append(
                    {
                        "transcription": result,
                        "boundaries": segment_boundaries,
                    }
                )