spec-to-wav synthesis produces some errors from existing mels

[roedoejet]

For some reason I had to change write(f"{data_path}.wav", sr, wav) to write(f"{data_path}.wav", sr, wav[0]) - I should investigate in hfgl/cli.py

[roedoejet]

related to #507

[SamuelLarkin]

I also stumbled on this using

everyvoice synthesize from-spec \
  --input preprocessed/spec/LJ033-0048--speaker_0--eng--spec-22050-mel-librosa.pt \
  --model logs_and_checkpoints/VocoderExperiment/base/checkpoints/last.ckpt

╭───────────────────────────────────────────────────── Traceback (most recent call last) ──────────────────────────────────────────────────────╮
│ /fs/hestia_Hnrc/ict/sam037/git/EveryVoice/everyvoice/model/vocoder/HiFiGAN_iSTFT_lightning/hfgl/cli.py:154 in synthesize                     │
│                                                                                                                                              │
│   151 │   except (TypeError, ValidationError) as e:                                                                                          │
│   152 │   │   logger.error(f"Unable to load {generator_path}: {e}")                                                                          │
│   153 │   │   sys.exit(1)                                                                                                                    │
│ ❱ 154 │   wav, sr = synthesize_data(data, vocoder_model, vocoder_config)                                                                     │
│   155 │   logger.info(f"Writing file {data_path}.wav")                                                                                       │
│   156 │   write(f"{data_path}.wav", sr, wav[0])                                                                                              │
│   157                                                                                                                                        │
│                                                                                                                                              │
│ /fs/hestia_Hnrc/ict/sam037/git/EveryVoice/everyvoice/model/vocoder/HiFiGAN_iSTFT_lightning/hfgl/utils.py:85 in synthesize_data               │
│                                                                                                                                              │
│    82 │   │   wavs = inverse_spectral_transform(mag * torch.exp(phase * 1j)).unsqueeze(-2)                                                   │
│    83 │   else:                                                                                                                              │
│    84 │   │   with torch.no_grad():                                                                                                          │
│ ❱  85 │   │   │   wavs = model.generator(data.transpose(1, 2))                                                                               │
│    86 │   # squeeze to remove the channel dimension                                                                                          │
│    87 │   return (                                                                                                                           │
│    88 │   │   wavs.squeeze(1).cpu().numpy(),                                                                                                 │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
IndexError: Dimension out of range (expected to be in range of [-2, 1], but got 2)