digiphyte/swivuriso-turbo

Swivuriso (turbo)

Swivuriso is a multilingual South African speech-recognition model covering seven South African languages: isiZulu, isiXhosa, Sesotho, Setswana, Xitsonga, isiNdebele, Tshivenda.

This is not our model. It is dsfsi-anv/za-anv-multilingual-whisper-v3-turbo by the Data Science for Social Impact (DSFSI) research group at the University of Pretoria, built as part of the African Next Voices project and named for the Swivuriso dataset it is trained on. It is MIT-licensed. We host it here, under its own name, only because we converted it to CTranslate2 (int8) so that faster-whisper can load it. No retraining was done. Full credit to DSFSI / African Next Voices (see NOTICE).

Use (faster-whisper)

from faster_whisper import WhisperModel
model = WhisperModel("digiphyte/swivuriso-turbo", device="cuda", compute_type="int8_float16")  # CPU: device="cpu", compute_type="int8"
# faster-whisper has no language codes for these seven languages, and the source model
# forces none, so run on auto-detect: pass language=None (do NOT force a language).
segments, info = model.transcribe("audio.wav", language=None, beam_size=5)
for s in segments:
    print(s.text)

Note on language=None: faster-whisper's language set does not include these seven languages, so you cannot (and should not) force one. The model decodes them correctly on auto-detect. faster-whisper may report the detected language as something else (e.g. "en") because it has no code to name them, but the transcribed text is the correct South African language. Use this model only for the seven languages it covers; for Afrikaans or English, use a Whisper or Fluister model instead.

Status

Beta. Transcription quality on real-world South African audio is still being verified.

Licence and attribution

MIT (see LICENSE), preserving the copyright of the model's authors. This repository is a format conversion of a derivative work; the model and its fine-tune (DSFSI / African Next Voices, MIT), the Swivuriso training data (African Next Voices), and the base model (OpenAI Whisper, Apache-2.0) are credited in NOTICE.