The following models are available:

• hparams/hparams_g2p_rnn.yaml: RNN-based (LSTM-encoder, GRU-decoder), with an attention mechanism Transformer
• hparams/hparams_g2p_transformer.yaml: Transformer/Conformer model

The datasets used here are available at the following locations:

• LibriG2P (no stress markers): https://huggingface.co/datasets/flexthink/librig2p-nostress
• LibriG2P (no stress markers, spaces preserved, with homographs): https://huggingface.co/datasets/flexthink/librig2p-nostress-space

The datasets are derived from the LibriSpeech-Alignments dataset (https://zenodo.org/record/2619474#.YbwRGi_73JM), optimized for training G2P models.

Decoding is performed with a beamsearch, optionally enhanced with language models.

Before proceeding, ensure you have installed the necessary additional dependencies. To do this, simply run the following command in your terminal:

pip install -r extra_requirements.txt

To run this recipe, do the following:

python train.py <hyperparameter file>

Example:

python train.py hparams/hparams_g2p_transformer.yaml

With the default hyperparameters, the system employs an LSTM encoder. The decoder is based on a standard GRU. The neural network is trained with negative-log.

With the default hyperparameters, the system employs a Conformer architecture with a convolutional encoder and a standard Transformer decoder.

The choice of Conformer vs Transformer is controlled by the transformer_encoder_module parameter.

Both RNN-based and Transformer-based models are capable of sentence-level hyperparameter disambiguation. Fine-tuning on homographs relies on an additional weighted loss computed only on the homograph and, therefore, requires a dataset in which the homograph is labelled, such as LibriG2P.

The experiment file is flexible enough to support a large variety of different systems. By properly changing the parameter files, you can try different encoders, decoders, and many other possible variations.

The provided G2P model allows for the optional use of a language model trained on the phoneme space integrated with a beam search in order to obtain a modest improvement in the results.

To train a language model, use the train_lm.py script provided.

For an RNN-based language model:

python train_lm.py hparams/hparams_lm_rnn.yaml

For a transformer-based language model:

python train_lm.py hparams/hparams_lm_transformer.yaml

To use a language model during training or inference

• Copy the language model file from the location indicated by <save_folder>/<checkpoint>/model.ckpt to <pretrained_path>/lm.ckpt
• Add--use_language_model true to the command line.

This recipe supports hyperparameter optimization via Oríon or other similar tools. For details on how to set up hyperparameter optimization, refer to the "Hyperparameter Optimization" tutorial in the Advanced Tutorials section on the SpeechBrain website:

https://speechbrain.github.io/tutorial_advanced.html

A supplemental hyperparameter file is provided for hyperparameter optimization, which will turn off checkpointing and limit the number of epochs:

hparams/hpopt.yaml

You can download LibriSpeech at http://www.openslr.org/12

| Release | hyperparams file | Sentence Test PER | Homograph % | Model link | |:-------------:|:--------------------------:| --------:| --------------------------------------------------------------------------------------------------:| | 0.5.12 | train_g2p_rnn.yaml | 2.72 | 94% | https://www.dropbox.com/sh/qmcl1obp8pxqaap/AAC3yXvjkfJ3mL-RKyAUxPdNa?dl=0 | | 0.5.12 | train_g2p_transformer.yaml | 2.89 | 92% | https://www.dropbox.com/sh/zhrxg7anuhje7e8/AADTeJtdsja_wClkE2DsF9Ewa?dl=0 |

NOTE: Sentence PER is reported as achieved at the end of the sentence training step. Nominal PER on librispeech data may increase post fine-tuning due to a distribution shift in labeling, if reevaluated. To replicate the result exactly, train with --homograph_epochs=0.

Pretrained language models can be found at the following URLs:

• RNN: https://www.dropbox.com/sh/pig0uk80xxii7cg/AACQ1rrRLYthvpNZ5FadPLtRa?dl=0
• Transformer: https://www.dropbox.com/sh/tkf6di10edpz4i6/AAArnGAkE0bEEOvOGfc6KWuma?dl=0

The best model is available on HuggingFace: https://huggingface.co/speechbrain/soundchoice-g2p

All reference times are given for a Quattro P5000 GPU. These are rough estimations only - exact training times will vary depending on the hyperparameters chosen and system configuration.

• Lexicon: approx. 6 minutes/epoch for training, < 1min/epoch for evaluation
• Sentence: approx. 50 minutes/epoch for training, 55 minutes/epoch for evaluation

• Lexicon: approx. 7 minutes/epoch for training, approx. 1/2 minutes/epoch for evaluation
• Sentence: 25-30 minutes/epoch for training, ~1 hour and 45 minutes/epoch for evaluation

Note: To speed up evaluation with the Transformer model, consider reducing the beam size. For fastest evaluation, use --beam_search_beam_size 1.

Pretrained models can be found on the following Google drive: https://www.dropbox.com/sh/3m4u7xda4xsh2ob/AAAYpOJHRhYbUHmuQtybgzrea?dl=0

• Website: https://speechbrain.github.io/
• Code: https://github.com/speechbrain/speechbrain/
• HuggingFace: https://huggingface.co/speechbrain/

Please, cite SpeechBrain if you use it for your research or business.

@misc{speechbrainV1,
  title={Open-Source Conversational AI with SpeechBrain 1.0},
  author={Mirco Ravanelli and Titouan Parcollet and Adel Moumen and Sylvain de Langen and Cem Subakan and Peter Plantinga and Yingzhi Wang and Pooneh Mousavi and Luca Della Libera and Artem Ploujnikov and Francesco Paissan and Davide Borra and Salah Zaiem and Zeyu Zhao and Shucong Zhang and Georgios Karakasidis and Sung-Lin Yeh and Pierre Champion and Aku Rouhe and Rudolf Braun and Florian Mai and Juan Zuluaga-Gomez and Seyed Mahed Mousavi and Andreas Nautsch and Xuechen Liu and Sangeet Sagar and Jarod Duret and Salima Mdhaffar and Gaelle Laperriere and Mickael Rouvier and Renato De Mori and Yannick Esteve},
  year={2024},
  eprint={2407.00463},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2407.00463},
}
@misc{speechbrain,
  title={{SpeechBrain}: A General-Purpose Speech Toolkit},
  author={Mirco Ravanelli and Titouan Parcollet and Peter Plantinga and Aku Rouhe and Samuele Cornell and Loren Lugosch and Cem Subakan and Nauman Dawalatabad and Abdelwahab Heba and Jianyuan Zhong and Ju-Chieh Chou and Sung-Lin Yeh and Szu-Wei Fu and Chien-Feng Liao and Elena Rastorgueva and François Grondin and William Aris and Hwidong Na and Yan Gao and Renato De Mori and Yoshua Bengio},
  year={2021},
  eprint={2106.04624},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  note={arXiv:2106.04624}
}