faster-whisper-51-african-languages

CTranslate2 (float16) conversion of Sunbird/SunflowerASR-51-african-languages — Whisper large-v3 fine-tuned for ASR in 51 African languages — for use with faster-whisper.

Converted from source revision 9177ad6c.

A different compute_type (e.g. int8_float16, ~1.6 GB VRAM) can be selected at load time and is quantized on the fly.

IMPORTANT: language codes

The fine-tune reuses Whisper's built-in language-token slots for African languages, and faster-whisper only accepts standard Whisper language codes. To transcribe a given language, pass the faster-whisper code from the table below (also machine-readable in language_map.json). Do not rely on automatic language detection.

Usage

import json
from faster_whisper import WhisperModel
from huggingface_hub import hf_hub_download

repo = "Sunbird/faster-whisper-51-african-languages"
lang_map = json.load(open(hf_hub_download(repo, "language_map.json")))

model = WhisperModel(repo, device="cuda", compute_type="float16")

segments, info = model.transcribe(
    "audio.wav",
    language=lang_map["lug"],   # Luganda
    task="transcribe",
    beam_size=5,
    vad_filter=True,
    condition_on_previous_text=False,
)
for segment in segments:
    print(f"[{segment.start:.2f} -> {segment.end:.2f}] {segment.text}")

Notes

• The source repo's reference pipeline normalizes input features (do_normalize=True in WhisperProcessor); faster-whisper's feature extractor does not apply this normalization. Validate WER on your target languages against the Transformers baseline before production use.
• See the base model card for training data, per-language WER/CER, and evaluation methodology.