sulaimank/wav2vec2-xlsr-CV_Fleurs_AMMI_ALFFA-swahili-200hrs

wav2vec2-xls-r-300m — Swahili ASR (200 hours)

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m for Swahili automatic speech recognition. It was originally trained by the ASR Africa research team and is hosted here for use as a base model for further fine-tuning on Luganda and other Bantu languages.

Model Performance

Evaluated on a combined held-out test set from CV + Fleurs + AMMI + ALFFA (Swahili):

Metric	Score
WER	13.73%
CER	4.54%

Training Hyperparameters

• Learning rate: 3e-4
• Train batch size: 8 (effective: 16 with gradient accumulation)
• Epochs: 100
• Optimizer: AdamW
• Mixed precision: FP16

How to Use

import torch
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

model_id = "sulaimank/wav2vec2-xlsr-CV_Fleurs_AMMI_ALFFA-swahili-200hrs"
processor = Wav2Vec2Processor.from_pretrained(model_id)
model     = Wav2Vec2ForCTC.from_pretrained(model_id)

# load your audio at 16kHz
input_values = processor(audio_array, sampling_rate=16000, return_tensors="pt").input_values

with torch.no_grad():
    logits = model(input_values).logits

predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)
print(transcription)