Ozen Turbo Hebrew a3 (whisper.cpp q5_0, for Android)

Ozen's noise-trained Turbo Hebrew (run a3, step 20774), the same weights as model-ozen-turbo-hebrew-a3-8bit-1, in whisper.cpp's format for the Android app.

Converted with whisper.cpp v1.9.5's own converter and quantized to q5_0 with its quantize tool. On ivrit.ai's WhatsApp test set (the 36 shortest of its 54 voice messages, greedy decoding, on a PC processor) the 16-bit, q8_0 and q5_0 files got 10.5%, 10.7% and 10.6% of words wrong, so q5_0 costs nothing measurable.

• ggml-ozen-a3-q5_0.bin: 574,041,195 bytes, sha256 c74b66691f8f97b4a38d43b5d8ee4ec1b32b3951a99520250f36883822c910af

Ozen Turbo Hebrew a3 (8-bit)

Noise-trained Turbo Hebrew, run a3, step 20774.

Hebrew Whisper Turbo (ivrit.ai), 8-bit, model release 1

The Hebrew-trained Whisper for Ozen's on-device captions, as the files the app downloads (see scripts/model-release/README.md). Not an app build.

• Source checkpoint: ivrit-ai/whisper-large-v3-turbo (Apache-2.0), OpenAI's Whisper large-v3-turbo continued-trained by ivrit.ai on about 5,000 hours of Hebrew.
• Converted to WhisperKit's Core ML layout with argmaxinc/whisperkittools (SplitHeadsQ encoder attention, Cat decoder attention, token timestamps on), then every weight palettized to 8 bits (uniform), 819 MB in all. The packages are uncompiled; the phone compiles them on first use.
• Assets are the model folder flattened: manifest.json lists each file's path, size and SHA-256.