README.md

metadata

license: cc-by-4.0
language:
  - qu
metrics:
  - cer
  - wer
pipeline_tag: automatic-speech-recognition
datasets:
  - second_americas_nlp_2022
tags:
  - audio
  - automatic-speech-recognition
  - speech
model-index:
  - name: Wav2Vec2 XLSR 300M Quechua Model by M Romero and Ivan G Torre
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Americas NLP 2022 qu
          type: second_americas_nlp_2022
          args: qu
        metrics:
          - name: Test CER
            type: cer
            value: 11.11
          - name: Test WER
            type: wer
            value: 11.11

Usage

The model can be used directly (without a language model) as follows:

from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
import torch
import torchaudio

# load model and processor
processor = Wav2Vec2Processor.from_pretrained("ivangtorre/wav2vec2-xlsr-300m-quechua")
model = Wav2Vec2ForCTC.from_pretrained("ivangtorre/wav2vec2-xlsr-300m-quechua")

# load dummy dataset and read soundfiles
file = torchaudio.load("quechua000573.wav")

# retrieve logits
logits = model(file[0]).logits

# take argmax and decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)
print("HF prediction: ", transcription)

This code snipnet shows how to Evaluate the wav2vec2-xlsr-300m-quechua in Second Americas NLP 2022 Quechua dev set

from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import torch
from jiwer import cer
import torch.nn.functional as F


librispeech_eval = load_dataset("ivangtorre/second_americas_nlp_2022", split="validation")

model = Wav2Vec2ForCTC.from_pretrained("ivangtorre/wav2vec2-xlsr-300m-quechua")
processor = Wav2Vec2Processor.from_pretrained("ivangtorre/wav2vec2-xlsr-300m-quechua")

def map_to_pred(batch):
    wav = batch["audio"][0]["array"]
    feats = torch.from_numpy(wav).float()
    feats = F.layer_norm(feats, feats.shape) # Normalization performed during finetuning
    feats = torch.unsqueeze(feats, 0)
    logits = model(feats).logits
    predicted_ids = torch.argmax(logits, dim=-1)
    batch["transcription"] = processor.batch_decode(predicted_ids)
    return batch

result = librispeech_eval.map(map_to_pred, batched=True, batch_size=1)

print("CER:", cer(result["source_processed"], result["transcription"]))
print("WER:", cer(result["source_processed"], result["transcription"]))

Citation

@misc{grosman2021xlsr-1b-russian,
  title={Fine-tuned {XLS-R} 1{B} model for speech recognition in {R}ussian},
  author={Grosman, Jonatas},
  howpublished={\url{https://huggingface.co/jonatasgrosman/wav2vec2-xls-r-1b-russian}},
  year={2022}
}