pejmans21/whisper-large-v3-turbo-persian-ggml

Whisper large-v3-turbo — Persian, GGML q5_0

A q5_0 quantization of the GGML file in givia/whisper-large-v3-turbo-persian, for whisper.cpp. It is the speech model of Voice Writing, an open-source macOS dictation app.

• Source: ggml-persian-v3-turbo.bin at revision 9626d38558f04c02f480b702c9db740eea567911 (SHA-256 743c4030ecb1a1d56ae64ced7a1e20ae8e51b38a87e218e995ee590fa556fff1).
• Change: whisper-quantize <source> ggml-large-v3-turbo-persian-q5_0.bin q5_0. Nothing else.
• File: ggml-large-v3-turbo-persian-q5_0.bin, 574,041,195 bytes, SHA-256 d832fe5cd81c756bc1962765cff4a33cacde00a7701809837875245870b6d5a8.

Results

whisper.cpp, greedy decoding, temperature fallback off (-nf; with it on, this fine-tune re-decodes often and is about 3× slower and less accurate).

Model	Persian WER / CER (96 recordings, ~50 speakers)	English FLEURS WER
openai large-v3-turbo q5_0	42.1 / 15.4	4.8
this file	22.8 / 12.4	6.4

The Persian references were machine transcripts of mostly read speech, so treat the absolute numbers as rough.

License and attribution

CC BY 4.0, as the source model. "Whisper large-v3-turbo — Persian (Farsi)" by givia, itself a fine-tune of openai/whisper-large-v3-turbo (MIT). Changes: quantized to q5_0, as above.