Amharic ASR using fine-tuned Wav2vec2 XLSR-53

This is a finetuned version of facebook/wav2vec2-large-xlsr-53 trained on the Amharic Speech Corpus. This corpus was produced by Abate et al. (2005) (10.21437/Interspeech.2005-467).

The model achieves a WER of 26% and a CER of 7% on the validation set of the Amharic Readspeech data.

Usage

The model can be used as follows:

import librosa
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

model = Wav2Vec2ForCTC.from_pretrained("agkphysics/wav2vec2-large-xlsr-53-amharic")
processor = Wav2Vec2Processor.from_pretrained("agkphysics/wav2vec2-large-xlsr-53-amharic")

audio, _ = librosa.load("/path/to/audio.wav", sr=16000)

input_values = processor(
    audio.squeeze(),
    sampling_rate=16000,
    return_tensors="pt"
).input_values

model.eval()
with torch.no_grad():
    logits = model(input_values).logits
    preds = logits.argmax(-1)
    texts = processor.batch_decode(preds)
print(texts[0])

Training

The code to train this model is available at https://github.com/agkphysics/amharic-asr.