datasets:
- alaleye/aloresb
metrics:
- wer
pipeline_tag: automatic-speech-recognition
Wav2vec2-Bert-Fongbe
This model is a fine-tuned version of facebook/w2v-bert-2.0. This has a WER of 24.20 on Aloresb dataset, fongbe language.
Model description
This model is a fine-tuned version of the wav2vec2-BERT architecture on the AlorésB dataset for the Fongbe language. Fongbe, a Gbe language, is predominantly spoken in the southern region of Benin. The model has been fine-tuned specifically for Automatic Speech Recognition (ASR) tasks in this language. It can be useful for transcription services, research, and linguistic studies involving Fongbe.
Details
- Model Name: wav2vec2-bert-fongbe
- Base Model: facebook/w2v-bert-2.0
- Fine-tuned on: Aloresb dataset
- Languages: Fongbe
- Architecture: Wav2vec2 + BERT
- Fine-tuning Dataset: Aloresb (Fongbe)
How to use
import torch
import soundfile as sf
from transformers import AutoModelForCTC, Wav2Vec2BertProcessor
model_name = "OctaSpace/wav2vec2-bert-fongbe"
asr_model = AutoModelForCTC.from_pretrained(model_name).to(device)
processor = Wav2Vec2BertProcessor.from_pretrained(model_name)
audio_input, _ = sf.read(file)
inputs = processor([audio_input], sampling_rate=16_000).input_features
features = torch.tensor(inputs)
with torch.no_grad():
logits = asr_model(features).logits
predicted_ids = torch.argmax(logits, dim=-1)
predictions = processor.batch_decode(predicted_ids, skip_special_tokens=True)
Training Procedure
The model was fine-tuned on the Aloresb dataset, which contains audio recordings and transcriptions in Fongbe.
Training Parameters:
Optimizer: AdamW Learning Rate: 3e-5 Batch Size: 3 Epochs: 3 Evaluation Results The model was evaluated using the Word Error Rate (WER) metric on a test set. Here are the results:
WER: 24.20%