On-Device Speech AI · STT & Speaker Diarization · Senior Backend

Rust — on-device speech AI that runs fully offline after a one-time model download, shipped as a single binary: speech-to-text (gigastt) and speaker diarization (polyvoice) — the two projects I actively build — plus the zero-dependency WAVE codec underneath them (ryf). Local-first by default: no cloud APIs at runtime, no vendor lock-in. On the side — agent & dev tooling experiments in Rust / Go (see the fold below).

In private development — Russian TTS on CosyVoice3, punctuation & casing restoration for ASR output, English-to-Russian video dubbing, semantic search over call recordings.

Track record (Apr–Sep 2026, three speech repos) — ~1,400 commits · ~160k lines of Rust · ~3,000 tests · 100+ tagged releases · weekly WER / DER benchmarks in CI · issues and feature requests from external users.

Prior production backend — Node.js / TypeScript (NestJS, Moleculer) and Python (FastAPI, Django): REST APIs and gateways from scratch, message brokers, payment & telephony integrations, large legacy refactors, and email / document (PDF/DOCX) pipelines.

file RTF ~0.10 (≈10× real-time on M1 CPU) · 3.55% WER clean read (statistical tie vs Vosk) · phone calls 18.5% — beats Vosk and T-one · held-out Common Voice 2.63% — beats Vosk and faster-whisper · like-for-like 6-engine bench · one binary · no cloud at runtime

cargo install gigastt && gigastt serve
# WebSocket: ws://127.0.0.1:9876/v1/ws
# REST API:  http://127.0.0.1:9876/v1/transcribe

GigaAM v3 + ONNX Runtime · INT8 ~225 MB · opt-in multilingual heads (ru/en/kk/ky/uz) · WebSocket streaming with Silero VAD end-of-utterance · speaker diarization via polyvoice · REST / SSE + async jobs · optional punctuation, casing & ITN · Android FFI + NNAPI · CPU / CoreML / CUDA · cargo / brew / npm / pip / Docker · 64 releases, 1.3k tests

Who spoke when, on CPU, without Python · ~15% DER VoxConverse-test (collar 0; pyannote 3.1 = 11.3%) · ~160× real-time on a laptop CPU · ~8.4 MB MIT INT8 models · 7.5k+ crates.io downloads

cargo add polyvoice --features "pipeline-native,vbx"

Powerset segmentation + WeSpeaker embeddings + VBx (AHC / K-means / NME-SC alternatives) · overlap resegmentation · who-said-what via companion ASR crates · Rust / Python / C FFI / CLI / MCP · crates.io + PyPI

Powers the diarization feature in gigastt, phonex, nihostt and phostt.

• ryf — WAVE codec, zero default deps · PCM / IEEE float / G.711 / G.722 / GSM 06.10 / ADPCM · RIFF / RF64 / BW64 / Wave64 · ffmpeg as test oracle · WAV ingest for gigastt and polyvoice · crates.io
• phonex — multilingual on-device STT engine · offline + streaming, single binary, HTTP / gRPC / FFI · crates.io
• nihostt — Japanese STT server (ReazonSpeech-k2-v2) · WebSocket / REST / SSE · crates.io + Homebrew + Docker
• phostt — Vietnamese STT server (Zipformer-vi RNN-T) · streaming · crates.io + pypi
• localmt — offline Android translation SDK · GGUF / llama.cpp via JNI · pre-release