Al-Farabi Assistant — ASR and TTS installation bundle
Private distribution of the exact speech models used by the Al-Farabi Assistant. This is an installation archive with runtime code, tokenizers, voice prompts, source notices and resolved dependencies. It is not a new fine-tuned checkpoint or a directly loadable Transformers repository.
| Component | Repository | Pinned revision |
|---|---|---|
| ASR | nur-dev/asr-220m-kk-ru-en |
45577ca31f29eb2ad5e548f45ab157123e1e863a |
| TTS | nur-dev/ait-syn-omnivoice-72-voices |
6f555261fc167ce98f994b78c9441af2eda2300c |
The default voice is kk_03, speed 1.0. ASR accepts 16 kHz mono PCM; TTS produces 24 kHz mono PCM after each complete phrase. All 72 native-language voice presets are retained: 16 Kazakh, 24 Russian and 32 English. Russian/English use of the selected Kazakh kk_03 prompt remains experimental.
Files and integrity
farabi-speech-models.zip contains 395 payload files, including model weights and runtime assets. Its size is 3,434,616,968 bytes. The ZIP uses stored entries/ZIP64 to preserve weights without expensive recompression.
SHA-256:
639921a0c768d743b33031589f6bf6310a617022e1ef8fd3b835a9a5bc138525
The matching .sha256 file verifies transfer integrity. speech-models.manifest.json records every payload file's size and SHA-256, exact releases and platform defaults. The installer rejects unsafe/duplicate ZIP paths, checksum errors, changed checkpoint identity and conflicting existing files. Identical installed files are reused.
Install into the platform
Authenticate an authorized Hugging Face account during provisioning, then download this private repository. From the Al-Farabi Assistant project root:
hf download farabi-lab/farabi-speech-models --local-dir artifacts/speech-bundle
python3 artifacts/speech-bundle/ops/speech_models.py verify artifacts/speech-bundle/farabi-speech-models.zip
python3 artifacts/speech-bundle/ops/speech_models.py install artifacts/speech-bundle/farabi-speech-models.zip --root .
./farabi-app --open
The platform defaults select models/asr-220m-kk-ru-en and models/ait-syn-omnivoice-72-voices. Explicit service asr_model_dir and tts_model_dir settings can relocate them. Repeat installation after interruption; committed compatible files are not rewritten.
The ZIP does not include virtual environments, Python wheels, CUDA drivers, Qwen dialogue weights, the knowledge database or credentials. Provision Python 3.12, compatible CUDA and separate ASR/TTS environments using the included dependency locks before disconnected runtime. Downloading requires repository access; installation and model inference subsequently use local assets.
The archive was freshly restored and functionally checked with network connections blocked: ASR transcribed a Kazakh recording, the formatter/decoder/VAD loaded, and TTS synthesized kk_03 speech. All 72 prepared voice prompts loaded. CPU checks establish portability/function, not CPU real-time performance; the live deployment uses CUDA. Application and bundle tests passed 84 checks when the archive was prepared.
Platform source and operations.
Origins and licenses
The ASR contains a locally adapted GigaAM Conformer-CTC acoustic model, multilingual BERT formatting and Silero VAD. The TTS packages unchanged pretrained OmniVoice and original Higgs codec decoding modules with locally prepared recorded-speaker prompts. Each component and reference corpus retains its original terms; see LICENSES.md and the complete notices inside the archive. Packaging does not relicense model weights.