Skip to content

Speech models

Choose a tag to compare

View all tags
@imomzoda-dev imomzoda-dev released this 04 Nov 11:26
· 6 commits to main since this release
models-1
0ddb04d

Speech models used by VoiceFlow AI. They are downloaded by the app (Settings → Recognition); you do not need to download them by hand.

ggml-tajik-small-q5_1.bin — Tajik (тоҷикӣ) · 190 MB

Whisper Small fine-tuned for Tajik, in whisper.cpp format (ggml, q5_1 quantisation).

Test: 40 recordings of Google FLEURS Tajik (dev) Word errors Written in Cyrillic
Whisper Large v3 Turbo (general model) 102% 28%
This model 5.2% 100%

FLEURS is read speech; everyday speech will have more errors than this test.

  • Original model: abduaziz/whisper-small-tajik, Apache License 2.0, fine-tuned from openai/whisper-small (MIT License) on Google FLEURS (CC BY 4.0).
  • Changes: converted from Hugging Face safetensors to ggml (same tensor mapping as whisper.cpp's convert-h5-to-ggml.py) and quantised to q5_1 with whisper-quantize. The weights were not otherwise changed.
  • SHA-256: 4052aa37f4479c5230530b8ba2a57fb187cffe14cbd024c0bbdbed573fb46fd5

ggml-persian-turbo-q5_0.bin — Persian (فارسی) · 574 MB

Whisper Large v3 Turbo fine-tuned for Persian on ~236 hours of openly licensed speech (conversational and read), with punctuation; whisper.cpp format, q5_0 quantisation.

Test: 40 phrases ≤ 10 s of Google FLEURS Persian (dev) Word errors
Whisper Large v3 Turbo (general model, Persian prompt) 20.5%
This model 6.1%
  • Original model: givia/whisper-large-v3-turbo-persian (ggml-persian-v3-turbo.bin), Creative Commons Attribution 4.0, fine-tuned from openai/whisper-large-v3-turbo (MIT License).
  • Changes: quantised from f16 to q5_0 with whisper-quantize; accuracy is unchanged (40 full FLEURS clips: f16 10.2%, q5_0 10.1% word errors). The weights were not otherwise changed.
  • SHA-256: d832fe5cd81c756bc1962765cff4a33cacde00a7701809837875245870b6d5a8
Assets 4
Loading