Spaces:
Running
on
Zero
Running
on
Zero
File size: 8,005 Bytes
45ee559 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 |
# Humble FAQ
We tried to collect common issues and questions we receive about 🐸TTS. It is worth checking before going deeper.
## Errors with a pre-trained model. How can I resolve this?
- Make sure you use the right commit version of 🐸TTS. Each pre-trained model has its corresponding version that needs to be used. It is defined on the model table.
- If it is still problematic, post your problem on [Discussions](https://github.com/coqui-ai/TTS/discussions). Please give as many details as possible (error message, your TTS version, your TTS model and config.json etc.)
- If you feel like it's a bug to be fixed, then prefer Github issues with the same level of scrutiny.
## What are the requirements of a good 🐸TTS dataset?
* {ref}`See this page <what_makes_a_good_dataset>`
## How should I choose the right model?
- First, train Tacotron. It is smaller and faster to experiment with. If it performs poorly, try Tacotron2.
- Tacotron models produce the most natural voice if your dataset is not too noisy.
- If both models do not perform well and especially the attention does not align, then try AlignTTS or GlowTTS.
- If you need faster models, consider SpeedySpeech, GlowTTS or AlignTTS. Keep in mind that SpeedySpeech requires a pre-trained Tacotron or Tacotron2 model to compute text-to-speech alignments.
## How can I train my own `tts` model?
0. Check your dataset with notebooks in [dataset_analysis](https://github.com/coqui-ai/TTS/tree/master/notebooks/dataset_analysis) folder. Use [this notebook](https://github.com/coqui-ai/TTS/blob/master/notebooks/dataset_analysis/CheckSpectrograms.ipynb) to find the right audio processing parameters. A better set of parameters results in a better audio synthesis.
1. Write your own dataset `formatter` in `datasets/formatters.py` or format your dataset as one of the supported datasets, like LJSpeech.
A `formatter` parses the metadata file and converts a list of training samples.
2. If you have a dataset with a different alphabet than English, you need to set your own character list in the ```config.json```.
- If you use phonemes for training and your language is supported [here](https://github.com/rhasspy/gruut#supported-languages), you don't need to set your character list.
- You can use `TTS/bin/find_unique_chars.py` to get characters used in your dataset.
3. Write your own text cleaner in ```utils.text.cleaners```. It is not always necessary, except when you have a different alphabet or language-specific requirements.
- A `cleaner` performs number and abbreviation expansion and text normalization. Basically, it converts the written text to its spoken format.
- If you go lazy, you can try using ```basic_cleaners```.
4. Fill in a ```config.json```. Go over each parameter one by one and consider it regarding the appended explanation.
- Check the `Coqpit` class created for your target model. Coqpit classes for `tts` models are under `TTS/tts/configs/`.
- You just need to define fields you need/want to change in your `config.json`. For the rest, their default values are used.
- 'sample_rate', 'phoneme_language' (if phoneme enabled), 'output_path', 'datasets', 'text_cleaner' are the fields you need to edit in most of the cases.
- Here is a sample `config.json` for training a `GlowTTS` network.
```json
{
"model": "glow_tts",
"batch_size": 32,
"eval_batch_size": 16,
"num_loader_workers": 4,
"num_eval_loader_workers": 4,
"run_eval": true,
"test_delay_epochs": -1,
"epochs": 1000,
"text_cleaner": "english_cleaners",
"use_phonemes": false,
"phoneme_language": "en-us",
"phoneme_cache_path": "phoneme_cache",
"print_step": 25,
"print_eval": true,
"mixed_precision": false,
"output_path": "recipes/ljspeech/glow_tts/",
"test_sentences": ["Test this sentence.", "This test sentence.", "Sentence this test."],
"datasets":[{"formatter": "ljspeech", "meta_file_train":"metadata.csv", "path": "recipes/ljspeech/LJSpeech-1.1/"}]
}
```
6. Train your model.
- SingleGPU training: ```CUDA_VISIBLE_DEVICES="0" python train_tts.py --config_path config.json```
- MultiGPU training: ```python3 -m trainer.distribute --gpus "0,1" --script TTS/bin/train_tts.py --config_path config.json```
**Note:** You can also train your model using pure 🐍 python. Check ```{eval-rst} :ref: 'tutorial_for_nervous_beginners'```.
## How can I train in a different language?
- Check steps 2, 3, 4, 5 above.
## How can I train multi-GPUs?
- Check step 5 above.
## How can I check model performance?
- You can inspect model training and performance using ```tensorboard```. It will show you loss, attention alignment, model output. Go with the order below to measure the model performance.
1. Check ground truth spectrograms. If they do not look as they are supposed to, then check audio processing parameters in ```config.json```.
2. Check train and eval losses and make sure that they all decrease smoothly in time.
3. Check model spectrograms. Especially, training outputs should look similar to ground truth spectrograms after ~10K iterations.
4. Your model would not work well at test time until the attention has a near diagonal alignment. This is the sublime art of TTS training.
- Attention should converge diagonally after ~50K iterations.
- If attention does not converge, the probabilities are;
- Your dataset is too noisy or small.
- Samples are too long.
- Batch size is too small (batch_size < 32 would be having a hard time converging)
- You can also try other attention algorithms like 'graves', 'bidirectional_decoder', 'forward_attn'.
- 'bidirectional_decoder' is your ultimate savior, but it trains 2x slower and demands 1.5x more GPU memory.
- You can also try the other models like AlignTTS or GlowTTS.
## How do I know when to stop training?
There is no single objective metric to decide the end of a training since the voice quality is a subjective matter.
In our model trainings, we follow these steps;
- Check test time audio outputs, if it does not improve more.
- Check test time attention maps, if they look clear and diagonal.
- Check validation loss, if it converged and smoothly went down or started to overfit going up.
- If the answer is YES for all of the above, then test the model with a set of complex sentences. For English, you can use the `TestAttention` notebook.
Keep in mind that the approach above only validates the model robustness. It is hard to estimate the voice quality without asking the actual people.
The best approach is to pick a set of promising models and run a Mean-Opinion-Score study asking actual people to score the models.
## My model does not learn. How can I debug?
- Go over the steps under "How can I check model performance?"
## Attention does not align. How can I make it work?
- Check the 4th step under "How can I check model performance?"
## How can I test a trained model?
- The best way is to use `tts` or `tts-server` commands. For details check {ref}`here <synthesizing_speech>`.
- If you need to code your own ```TTS.utils.synthesizer.Synthesizer``` class.
## My Tacotron model does not stop - I see "Decoder stopped with 'max_decoder_steps" - Stopnet does not work.
- In general, all of the above relates to the `stopnet`. It is the part of the model telling the `decoder` when to stop.
- In general, a poor `stopnet` relates to something else that is broken in your model or dataset. Especially the attention module.
- One common reason is the silent parts in the audio clips at the beginning and the ending. Check ```trim_db``` value in the config. You can find a better value for your dataset by using ```CheckSpectrogram``` notebook. If this value is too small, too much of the audio will be trimmed. If too big, then too much silence will remain. Both will curtail the `stopnet` performance.
|