kodavoice-singlish

KodaVoice Singlish — CTranslate2 (int8)

A CTranslate2 int8 conversion of mjwong/whisper-large-v3-turbo-singlish, packaged so it loads directly in faster-whisper. Used as the default dictation model in KodaVoice for Singlish / Singapore–Malaysia English.

Only the format changed (PyTorch → CTranslate2, int8 quantization) — the weights are the source fine-tune's, unmodified.

Usage

from faster_whisper import WhisperModel

model = WhisperModel("salvestr0/kodavoice-singlish", device="cpu", compute_type="int8")
segments, info = model.transcribe("audio.wav", beam_size=1)
for s in segments:
    print(s.text)

beam_size=1 (greedy) is the recommended mode — beam search does not improve WER on this fine-tune.

whisper.cpp (GGML)

For the whisper.cpp / Vulkan runtime, this repo also ships a q5_0 GGML build (ggml-kodavoice-singlish-q5_0.bin, ~547 MB — same size/quality tier as the stock ggml-large-v3-turbo-q5_0):

whisper-cli -m ggml-kodavoice-singlish-q5_0.bin -f audio.wav -l en

Scope & honest limits

The fine-tune's accuracy gains are validated on Singapore + Malaysia English / Singlish. It is not a validated fine-tune for Philippine, Indonesian, Thai, or Vietnamese English — treat those as vanilla Whisper large-v3-turbo behaviour.

Attribution & license

• Source fine-tune: mjwong/whisper-large-v3-turbo-singlish — MIT, © mjwong
• Base model: openai/whisper-large-v3-turbo — MIT
• This derivative is released under MIT. See NOTICE for full attribution.