wolof_asr

Wolof ASR (wav2vec2 CTC)

Fine-tuned wav2vec2 CTC model for transcribing Wolof speech to text.

Model details

• Architecture: wav2vec2 CTC (AutoModelForCTC)
• Task: Automatic speech recognition (ASR)
• Language: Wolof (wo)
• Sample rate: 16 kHz mono
• Output: Plain text transcription

Expected input format

Quick start

from transformers import AutoModelForCTC, AutoProcessor
import librosa

model_id = "Or4kool/wolof_asr"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCTC.from_pretrained(model_id)
model.eval()

audio, sr = librosa.load("sample.wav", sr=16000, mono=True)
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")

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

predicted_ids = torch.argmax(logits, dim=-1)
text = processor.batch_decode(predicted_ids)[0]
print(text)

Pipeline usage

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="Or4kool/wolof_asr",
)
print(asr("sample.wav")["text"])

API service (this repository)

This model powers the myagro-audio service. Example requests:

JSON (base64)

curl -X POST https://your-endpoint/transcribe \
  -H "Content-Type: application/json" \
  -d '{
    "audio_base64": "<base64-audio>",
    "language": "wol",
    "return_timestamps": false
  }'

Multipart file upload

curl -X POST https://your-endpoint/transcribe/upload \
  -F "[email protected]" \
  -F "language=wol" \
  -F "return_timestamps=true"

Batch (RunPod-friendly)

{
  "input": {
    "items": [
      {"id": "clip-1", "audio_base64": "..."},
      {"id": "clip-2", "audio_base64": "..."}
    ],
    "return_timestamps": false
  }
}

Response format

{
  "text": "transcribed wolof text",
  "language": "wol",
  "duration_seconds": 2.5,
  "request_id": "uuid",
  "inference_ms": 120.4,
  "words": [
    {"word": "example", "start": 0.12, "end": 0.48}
  ]
}

words is optional and only included when return_timestamps=true.

Limitations

• Best performance on clear, close-mic speech at 16 kHz
• Word timestamps are approximate CTC alignments, not forced alignments
• Not intended for real-time streaming transcription
• Long-form audio should be split into chunks under 60 seconds

Training data

Citation

@misc{or4kool-wolof-asr,
  author = {Samuel Oriaku},
  title = {Wolof ASR wav2vec2 CTC},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Or4kool/wolof_asr}}
}

Uploading this card to Hugging Face

huggingface-cli login
huggingface-cli upload Or4kool/wolof_asr huggingface_model_card/README.md README.md