| --- |
| license: mit |
| language: |
| - af |
| base_model: openai/whisper-tiny |
| library_name: ctranslate2 |
| pipeline_tag: automatic-speech-recognition |
| tags: |
| - whisper |
| - fluister |
| - afrikaans |
| - south-africa |
| - faster-whisper |
| - ctranslate2 |
| --- |
| |
| # Fluister (tiny) |
|
|
| **Fluister** is an Afrikaans-optimised Whisper. ("Fluister" is Afrikaans for "to whisper".) It is a |
| fine-tune of OpenAI Whisper (openai/whisper-tiny), merged into the base weights and converted to CTranslate2 |
| (int8) for use with faster-whisper. By DigiPhyte (Pty) Ltd, South Africa. |
|
|
| On Afrikaans audio it reduces Whisper's drift to Dutch-style spellings ("gebou" not "gebouw", "mense" not "mensen") compared to stock Whisper whisper-tiny. As one of the smallest Whisper sizes its overall accuracy is limited; please read Limitations below before using it. |
|
|
| ## Use (faster-whisper) |
|
|
| ```python |
| from faster_whisper import WhisperModel |
| model = WhisperModel("digiphyte/fluister-tiny", device="cuda", compute_type="int8_float16") # CPU: device="cpu", compute_type="int8" |
| segments, info = model.transcribe("audio.wav", language="af", beam_size=5) |
| for s in segments: |
| print(s.text) |
| ``` |
|
|
| Pass `language="af"`; the Fluister models are tuned for Afrikaans and SA English and should be told |
| the language rather than relying on auto-detect. |
|
|
| ## Limitations |
|
|
| Fluister narrows one specific failure: Whisper spelling Afrikaans as Dutch. It does not turn a small model into a large one. Absolute accuracy is still bounded by the base size, language auto-detect can still mislabel the audio (tell it `language="af"`), and proper nouns, numbers, and rare or technical terms can still be wrong. For English-only audio, stock Whisper or a larger size is usually the better choice. |
|
|
| **This is the smallest, fastest tier (tiny), meant for very modest or CPU-only machines.** It cuts the Dutch drift relative to stock Whisper whisper-tiny, but it is the least accurate model in the Fluister family: expect noticeably more errors overall, weaker Afrikaans/English code-switching (some Afrikaans words can leak into English passages), and weaker proper nouns. It is a speed-and-size trade-off, not a quality model. If your machine can run them, prefer `fluister-medium` or `fluister-large-v3`. |
|
|
| ## Licence and attribution |
|
|
| MIT (see `LICENSE`). This is a derivative work; the base model (OpenAI Whisper, Apache-2.0) and the training data (`andreoosthuizen/afrikaans-30s`, CC-BY-4.0) (this size was fine-tuned by DigiPhyte directly on that dataset) are credited in `NOTICE`. |
|
|
