Generated WAV garbled after 40 seconds

[Manamama]

Describe the bug

Generation of longer phrases seems to garble the output after about the 30 second timestamp. The resulting files sound random: one was 60 second long, while this attached lasts only 40 seconds.

My box:

Python 3.8.10
OS: Ubuntu 20.04.5 LTS x86_64 
Kernel: 5.4.0-125-generic 
CPU: Intel i5-8265U (8) @ 3.900GHz 
GPU: Intel UHD Graphics 620 
Memory: 5598MiB / 7843MiB 

To Reproduce

tts --text "Nowhere in those kerosene years could she find a soft-headed match. The wife crossed over an ocean, red-faced and cheerless. She traded the flat pad of a stethoscope for a dining hall spatula. Life is two choices, she thinks: you hatch a life, or you pass through one. Photographs of a child swaddled in layers arrived by post. Money didn’t, to her embarrassment." --out_path test.wav && mplayer test.wav test1.wav.tar.gz test3.wav.tar.gz

Expected behavior

Non-garbled WAV file

Logs

/usr/lib/python3/dist-packages/paramiko/transport.py:219: CryptographyDeprecationWarning: Blowfish has been deprecated
  "class": algorithms.Blowfish,
 > tts_models/en/ljspeech/tacotron2-DDC is already downloaded.
 > vocoder_models/en/ljspeech/hifigan_v2 is already downloaded.
 > Using model: Tacotron2
 > Setting up Audio Processor...
 | > sample_rate:22050
 | > resample:False
 | > num_mels:80
 | > log_func:np.log
 | > min_level_db:-100
 | > frame_shift_ms:None
 | > frame_length_ms:None
 | > ref_level_db:20
 | > fft_size:1024
 | > power:1.5
 | > preemphasis:0.0
 | > griffin_lim_iters:60
 | > signal_norm:False
 | > symmetric_norm:True
 | > mel_fmin:0
 | > mel_fmax:8000.0
 | > pitch_fmin:0.0
 | > pitch_fmax:640.0
 | > spec_gain:1.0
 | > stft_pad_mode:reflect
 | > max_norm:4.0
 | > clip_norm:True
 | > do_trim_silence:True
 | > trim_db:60
 | > do_sound_norm:False
 | > do_amp_to_db_linear:True
 | > do_amp_to_db_mel:True
 | > do_rms_norm:False
 | > db_level:None
 | > stats_path:None
 | > base:2.718281828459045
 | > hop_length:256
 | > win_length:1024
 > Model's reduction rate `r` is set to: 1
 > Vocoder Model: hifigan
 > Setting up Audio Processor...
 | > sample_rate:22050
 | > resample:False
 | > num_mels:80
 | > log_func:np.log
 | > min_level_db:-100
 | > frame_shift_ms:None
 | > frame_length_ms:None
 | > ref_level_db:20
 | > fft_size:1024
 | > power:1.5
 | > preemphasis:0.0
 | > griffin_lim_iters:60
 | > signal_norm:False
 | > symmetric_norm:True
 | > mel_fmin:0
 | > mel_fmax:8000.0
 | > pitch_fmin:0.0
 | > pitch_fmax:640.0
 | > spec_gain:1.0
 | > stft_pad_mode:reflect
 | > max_norm:4.0
 | > clip_norm:True
 | > do_trim_silence:False
 | > trim_db:60
 | > do_sound_norm:False
 | > do_amp_to_db_linear:True
 | > do_amp_to_db_mel:True
 | > do_rms_norm:False
 | > db_level:None
 | > stats_path:None
 | > base:2.718281828459045
 | > hop_length:256
 | > win_length:1024
 > Generator Model: hifigan_generator
 > Discriminator Model: hifigan_discriminator
Removing weight norm...
 > Text: Nowhere in those kerosene years could she find a soft-headed match. The wife crossed over an ocean, red-faced and cheerless. She traded the flat pad of a stethoscope for a dining hall spatula. Life is two choices, she thinks: you hatch a life, or you pass through one. Photographs of a child swaddled in layers arrived by post. Money didn’t, to her embarrassment.
 > Text splitted to sentences.
['Nowhere in those kerosene years could she find a soft-headed match.', 'The wife crossed over an ocean, red-faced and cheerless.', 'She traded the flat pad of a stethoscope for a dining hall spatula.', 'Life is two choices, she thinks: you hatch a life, or you pass through one.', 'Photographs of a child swaddled in layers arrived by post.', 'Money didn’t, to her embarrassment.']
money didn’t, to her embarrassment.
 [!] Character '’' not found in the vocabulary. Discarding it.
   > Decoder stopped with `max_decoder_steps` 10000
 > Processing time: 113.27926182746887
 > Real-time factor: 0.7940720796039662
 > Saving output to test.wav

### Environment

```shell
{
    "CUDA": {
        "GPU": [],
        "available": false,
        "version": "10.2"
    },
    "Packages": {
        "PyTorch_debug": false,
        "PyTorch_version": "1.12.1+cu102",
        "TTS": "0.8.0",
        "numpy": "1.21.6"
    },
    "System": {
        "OS": "Linux",
        "architecture": [
            "64bit",
            "ELF"
        ],
        "processor": "x86_64",
        "python": "3.8.10",
        "version": "#141-Ubuntu SMP Wed Aug 10 13:42:03 UTC 2022"
    }
}

Additional context

The generated WAV files have random length.

[thorstenMueller]

Imho this isn't a bug. You could increase the max_decoder_steps in config.json, but this might lead to problems synthesizing shorter phrases. Or split your text and run synthesizing for multiple shorter phrases.

[Manamama]

Thanks.

1. I could not find max_decoder_steps in the config.json located in the compiled: ~/.local/lib/python3.8/site-packages/TTS/server/conf.json nor in ~/.local/lib/python3.8/site-packages/TTS/tts I could not find it in the jsons in the original cloned files that quick, either. It also seems to be missing from the tts command line parameters, as well.

-> Where is that option located? (Do bear with me here, as I am not a developer or programmer.)

2. If "it's not a bug but a feature", don't the input get split automatically?: > Text splitted to sentences. to accommodate for this?

3. FYI, the garbled output happens also with other models, but does not happen that often with other sample longer texts; yet I have tested 5 samples so far only.

[Manamama]

Another bug, but most probably due to the paucity of the models themselves, the Dearest Creature in Creation poem is mispronounced, while e.g. Google's and Nuance TTS handle it fine.

[erogol]

I'd suggest trying other English models.the default model can sometimes garble.

Comparing this open source project with Google and Nuance is flattering but hard to match :)

I move this to discussions as it it not a dev issue.