license: apache-2.0
base_model: facebook/wav2vec2-large-xlsr-53
metrics:
- wer
model-index:
- name: wav2vec2-xlsr-53-ft-ccv-en-cy
results: []
datasets:
- techiaith/commonvoice_16_1_en_cy
language:
- cy
- en
pipeline_tag: automatic-speech-recognition
wav2vec2-xlsr-53-ft-cy-en-withlm
This model is a version of facebook/wav2vec2-large-xlsr-53 that has been fined-tuned with a custom bilingual datasets derived from the Welsh and English data releases of Mozilla Foundation's Commonvoice project. See : techiaith/commonvoice_16_1_en_cy.
In addition, this model also includes a single KenLM n-gram model trained with balanced collections of Welsh and English texts from OSCAR This avoids the need for any language detection for determining whether to use a Welsh or English n-gram models during CTC decoding.
Usage
The wav2vec2-xlsr-53-ft-cy-en-withlm model can be used directly as follows:
import torch
import torchaudio
import librosa
from transformers import Wav2Vec2ForCTC, Wav2Vec2ProcessorWithLM
processor = Wav2Vec2ProcessorWithLM.from_pretrained("techiaith/wav2vec2-xlsr-53-ft-cy-en-withlm")
model = Wav2Vec2ForCTC.from_pretrained("techiaith/wav2vec2-xlsr-53-ft-cy-en-withlm")
audio, rate = librosa.load(<path/to/audio_file>, sr=16000)
inputs = processor(audio, sampling_rate=16_000, return_tensors="pt", padding=True)
with torch.no_grad():
tlogits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
print("Prediction: ", processor.batch_decode(tlogits.numpy(), beam_width=10).text[0].strip())
Usage with a pipeline is even simpler...
from transformers import pipeline
transcriber = pipeline("automatic-speech-recognition", model="techiaith/wav2vec2-xlsr-53-ft-cy-en-withlm")
def transcribe(audio):
return transcriber(audio)["text"]
transcribe(<path/or/url/to/any/audiofile>)
Evaluation
According to a balanced English+Welsh test set derived from Common Voice version 16.1, the WER of techiaith/wav2vec2-xlsr-53-ft-cy-en-withlm is 23.79%
However, when evaluated with language specific test sets, the model exhibits a bias to perform better with Welsh.
| Common Voice Test Set Language | WER | CER |
|---|---|---|
| EN+CY | 23.79 | 9.68 |
| EN | 34.47 | 14.83 |
| CY | 12.34 | 3.55 |
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0003
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 64
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 800
- training_steps: 9000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 6.0574 | 0.25 | 500 | 2.0297 | 0.9991 |
| 1.224 | 0.5 | 1000 | 0.5368 | 0.4342 |
| 0.434 | 0.75 | 1500 | 0.4861 | 0.3891 |
| 0.3295 | 1.01 | 2000 | 0.4301 | 0.3411 |
| 0.2739 | 1.26 | 2500 | 0.3818 | 0.3053 |
| 0.2619 | 1.51 | 3000 | 0.3894 | 0.3060 |
| 0.2517 | 1.76 | 3500 | 0.3497 | 0.2802 |
| 0.2244 | 2.01 | 4000 | 0.3519 | 0.2792 |
| 0.1854 | 2.26 | 4500 | 0.3376 | 0.2718 |
| 0.1779 | 2.51 | 5000 | 0.3206 | 0.2520 |
| 0.1749 | 2.77 | 5500 | 0.3169 | 0.2535 |
| 0.1636 | 3.02 | 6000 | 0.3122 | 0.2465 |
| 0.137 | 3.27 | 6500 | 0.3054 | 0.2382 |
| 0.1311 | 3.52 | 7000 | 0.2956 | 0.2280 |
| 0.1261 | 3.77 | 7500 | 0.2898 | 0.2236 |
| 0.1187 | 4.02 | 8000 | 0.2847 | 0.2176 |
| 0.1011 | 4.27 | 8500 | 0.2763 | 0.2124 |
| 0.0981 | 4.52 | 9000 | 0.2754 | 0.2115 |
Framework versions
- Transformers 4.38.2
- Pytorch 2.2.1+cu121
- Datasets 2.18.0
- Tokenizers 0.15.2