Fine-tuned wav2vec2-FR-7K-large model for ASR in French
This model is a fine-tuned version of LeBenchmark/wav2vec2-FR-7K-large, trained on a composite dataset comprising of over 2200 hours of French speech audio, using the train and validation splits of Common Voice 11.0, Multilingual LibriSpeech, Voxpopuli, Multilingual TEDx, MediaSpeech, and African Accented French. When using the model make sure that your speech input is also sampled at 16Khz.
Usage
- To use on a local audio file with the language model
import torch
import torchaudio
from transformers import AutoModelForCTC, Wav2Vec2ProcessorWithLM
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = AutoModelForCTC.from_pretrained("bhuang/asr-wav2vec2-french").to(device)
processor_with_lm = Wav2Vec2ProcessorWithLM.from_pretrained("bhuang/asr-wav2vec2-french")
model_sample_rate = processor_with_lm.feature_extractor.sampling_rate
wav_path = "example.wav" # path to your audio file
waveform, sample_rate = torchaudio.load(wav_path)
waveform = waveform.squeeze(axis=0) # mono
# resample
if sample_rate != model_sample_rate:
resampler = torchaudio.transforms.Resample(sample_rate, model_sample_rate)
waveform = resampler(waveform)
# normalize
input_dict = processor_with_lm(waveform, sampling_rate=model_sample_rate, return_tensors="pt")
with torch.inference_mode():
logits = model(input_dict.input_values.to(device)).logits
predicted_sentence = processor_with_lm.batch_decode(logits.cpu().numpy()).text[0]
- To use on a local audio file without the language model
import torch
import torchaudio
from transformers import AutoModelForCTC, Wav2Vec2Processor
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = AutoModelForCTC.from_pretrained("bhuang/asr-wav2vec2-french").to(device)
processor = Wav2Vec2Processor.from_pretrained("bhuang/asr-wav2vec2-french")
model_sample_rate = processor.feature_extractor.sampling_rate
wav_path = "example.wav" # path to your audio file
waveform, sample_rate = torchaudio.load(wav_path)
waveform = waveform.squeeze(axis=0) # mono
# resample
if sample_rate != model_sample_rate:
resampler = torchaudio.transforms.Resample(sample_rate, model_sample_rate)
waveform = resampler(waveform)
# normalize
input_dict = processor(waveform, sampling_rate=model_sample_rate, return_tensors="pt")
with torch.inference_mode():
logits = model(input_dict.input_values.to(device)).logits
# decode
predicted_ids = torch.argmax(logits, dim=-1)
predicted_sentence = processor.batch_decode(predicted_ids)[0]
Evaluation
- To evaluate on
mozilla-foundation/common_voice_11_0
python eval.py \
--model_id "bhuang/asr-wav2vec2-french" \
--dataset "mozilla-foundation/common_voice_11_0" \
--config "fr" \
--split "test" \
--log_outputs \
--outdir "outputs/results_mozilla-foundatio_common_voice_11_0_with_lm"
- To evaluate on
speech-recognition-community-v2/dev_data
python eval.py \
--model_id "bhuang/asr-wav2vec2-french" \
--dataset "speech-recognition-community-v2/dev_data" \
--config "fr" \
--split "validation" \
--chunk_length_s 30.0 \
--stride_length_s 5.0 \
--log_outputs \
--outdir "outputs/results_speech-recognition-community-v2_dev_data_with_lm"
- Downloads last month
- 1,086
This model can be loaded on the Inference API on-demand.
Datasets used to train bofenghuang/asr-wav2vec2-ctc-french
Space using bofenghuang/asr-wav2vec2-ctc-french 1
Evaluation results
- Test WER on Common Voice 11.0self-reported11.440
- Test WER (+LM) on Common Voice 11.0self-reported9.660
- Test WER on Multilingual LibriSpeech (MLS)self-reported5.930
- Test WER (+LM) on Multilingual LibriSpeech (MLS)self-reported5.130
- Test WER on VoxPopuliself-reported9.330
- Test WER (+LM) on VoxPopuliself-reported8.510
- Test WER on African Accented Frenchself-reported16.220
- Test WER (+LM) on African Accented Frenchself-reported15.390
- Test WER on Robust Speech Event - Dev Dataself-reported16.560
- Test WER (+LM) on Robust Speech Event - Dev Dataself-reported12.960
- Test WER on Fleursself-reported10.100
- Test WER (+LM) on Fleursself-reported8.840