Gigaam Ctc With Lm

🚀 GigaAM-v2-CTC with ngram LM and beamsearch 🤖 Hugging Face transformers

Russian ASR model GigaAM-v2-CTC with external ngram LM and beamsearch decoding, which can solve the problem of automatic speech recognition in Russian and provide accurate transcription results.

🚀 Quick Start

This is a Russian ASR (Automatic Speech Recognition) model named GigaAM-v2-CTC. It comes with an external n - gram LM (Language Model) and uses beamsearch decoding. You can use it just like other transformers ASR models.

✨ Features

• It is an original GigaAM - v2 - CTC with a transformers library interface.
• Supports beamsearch decoding and hypothesis rescoring with an external ngram LM.
• Can be used to extract word - level timestamps.

📦 Installation

My library versions:

• torch 2.5.1
• torchaudio 2.5.1
• transformers 4.49.0

You need to install kenlm and pyctcdecode:

pip install kenlm
pip install pyctcdecode

💻 Usage Examples

Basic Usage

from transformers import AutoModel, AutoProcessor
import torch
import torchaudio

# load audio
wav, sr = torchaudio.load("audio.wav")
# resample if necessary
wav = torchaudio.functional.resample(wav, sr, 16000)

# load model and processor
processor = AutoProcessor.from_pretrained("waveletdeboshir/gigaam-ctc-with-lm", trust_remote_code=True)
model = AutoModel.from_pretrained("waveletdeboshir/gigaam-ctc-with-lm", trust_remote_code=True)
model.eval()

input_features = processor(wav[0], sampling_rate=16000, return_tensors="pt")

# predict
with torch.no_grad():
    logits = model(**input_features).logits

# decoding with beamseach and LM (tune alpha, beta, beam_width for your data)
transcription = processor.batch_decode(
    logits=logits.numpy(),
    beam_width=64,
    alpha=0.5,
    beta=0.5,
).text[0]

Advanced Usage

We can use the decoder to extract word - level timestamps. For this, we need to know the model stride and set the parameter output_word_offsets=True. In our case (Conformer), MODEL_STRIDE = 40 ms per timestamp.

MODEL_STRIDE = 40
outputs = processor.batch_decode(
    logits=logits.numpy(),
    beam_width=64,
    alpha=0.5,
    beta=0.5,
    output_word_offsets=True
)
word_ts = [
    {
        "word": d["word"],
        "start": round(d["start_offset"] * MODEL_STRIDE / 1000, 2),
        "end": round(d["end_offset"] * MODEL_STRIDE / 1000, 2),
    }
    for d in outputs.word_offsets[0]
]

📚 Documentation

• Original git: https://github.com/salute-developers/GigaAM
• ngram LM source: bond005/wav2vec2-large-ru-golos-with-lm
• The file gigaam_transformers.py contains model, feature extractor and tokenizer classes with usual transformers methods. The model can be initialized with transformers auto classes.

📄 License

This project is licensed under the MIT license.