Voxcribe models
whisper.cpp (ggml) conversions of
GoranS/whisper-large-v3-turbo-hr-parla
(revision 3247238374e3d81f55b1451a105294b306e093bd), a Croatian fine-tune of
OpenAI Whisper large-v3-turbo by GoranS, used by the Voxcribe offline transcription app.
The weights are not modified apart from the format conversion and, for the q5_0 file, quantization.
| File | Format | Size (bytes) | SHA-256 |
|---|---|---|---|
ggml-goran-turbo-hr-f16.bin |
f16 | 1624555275 | e5c24f659ff835caff38e194a95f74de34e1636866b2158f878113f41dddf99b |
ggml-goran-turbo-hr-q5_0.bin |
q5_0 | 574041195 | f67c099966bea9fa27c0f092e3fd5b82d3933095095037d3c4190f3d216fb8c7 |
The tokenizer and mel filters are those of ggml-large-v3-turbo from
ggerganov/whisper.cpp (the fine-tune keeps the vocabulary).
Usage note
The fine-tune was not trained on timestamps: on long audio whisper.cpp skips large parts of each 30 s window. Transcribe short speech segments (up to ~8 s, e.g. cut with Silero VAD) instead.
On 15 FLEURS test chunks (read speech, segmented at pauses) the word error rate was:
| Model | Croatian | Bosnian |
|---|---|---|
| whisper large-v3-turbo q5_0, long-form | 18.0 % | 14.3 % |
| this model, f16, ≤ 8 s segments | 11.5 % | 12.5 % |
License
Apache License 2.0, as the base model. Original model: © GoranS; based on OpenAI Whisper large-v3-turbo (MIT).
Model tree for ldanijel/voxcribe-models
Base model
openai/whisper-large-v3