README.md

TranScorerLM

Transformer as Scorer (Language Model) for STT accoustic models.

Get started

# Install with pip
❯ pip install git+https://github.com/wasertech/TranScorerLM.git

# Use TranScorer to convert an accoustic representation to text
❯ transcorer -f 'audio.wav'
Some weights of Wav2Vec2ForCTC were not initialized from the model checkpoint at facebook/wav2vec2-base-960h and are newly initialized: ['wav2vec2.masked_spec_embed']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
Loading TranScorer......Took 2.2624789050023537 second(s).
Loading audio.wav......Took 0.00048493899521417916 second(s).
Tokenizing......Took 0.0008750690030865371 second(s).
Decoding speech......Took 0.21528533000673633 second(s).
CAN I TEST YOU

When --file is a valid audio file, it is used to get the accoustic representation.

You can also use the scorer python module.

from pathlib import Path

from scorer.transformer import TranScorer

ts = TranScorer("path/to/scorer.transform")

to_transcribe = "path/to/audio.wav" 
transcript = ts.transcribe(to_transcribe)
print(f"{to_transcribe=} -> {transcript=}")

You could also ask the scorer to tokenize something for you.

python -m scorer.transformer.tokenizer \
    --lang "english" \
    --data_path "parent/path/to/tokenizer/data/*.txt" \
    --test "This test sentence will be tokenized."

Training a new scorer from facebook/wav2vec2-large-xlsr-53

Start training a scorer using trainscorer.

# trainscorer -> python -m scorer.train
trainscorer \
    --model_name_or_path facebook/wav2vec2-base \
    --dataset_name wav2txt \
    --dataset_config_name wav2txt \
    --output_dir wav2txt/models \
    --overwrite_output_dir \
    --remove_unused_columns False \
    --do_train y \
    --do_eval y \
    --fp16 \
    --learning_rate 3e-5 \
    --max_length_seconds 1 \
    --attention_mask False \
    --warmup_ratio 0.1 \
    --num_train_epochs 5 \
    --per_device_train_batch_size 32 \
    --gradient_accumulation_steps 4 \
    --per_device_eval_batch_size 32 \
    --dataloader_num_workers 4 \
    --logging_strategy steps \
    --logging_steps 10 \
    --evaluation_strategy epoch \
    --save_strategy epoch \
    --load_best_model_at_end True \
    --metric_for_best_model accuracy \
    --save_total_limit 3 \
    --seed 0

You can also pass a --push_to_hub flag along with --hub_token '${HUB_API_TOKEN}' --push_to_hub_model_id '${MODEL_ID}' --use_auth_token y; where $HUB_API_TOKEN is a valid HuggingFace API Token (with write access) and $MODEL_ID is the name of the repo you want to create for your model, this will publish it on the hub under the above mentionned repository.

You can also train directly from a python script.

from scorer.train import train

if __name__ == "__main__":
    try:
        train()
        sys.exit(0)
    except KeyboardInterrupt:
        sys.exit(1)

Training will create a custom tokenizer using availible sentences.

License

This project is distributed under Mozilla Public License 2.0.

Contribute

Please read CONTRIBUTE before anthing.