ekhodzitsky/gigastt
ekhodzitsky/gigastt has 84 GitHub stars and 11 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-16 19:11 UTC. Last push 2026-11-11.
Measured star growth
- in 24h
- 0
- in 7 days
- Update delayed
- in 15 days
- Update delayed
- in 1 month
- Update delayed
Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.
About
Local STT server powered by GigaAM v3.
SecurityREADME.md
Originalgigastt
Local speech recognition for files, live audio and embedded applications.
gigastt is a Rust CLI and HTTP/WebSocket server powered by GigaAM and ONNX Runtime. Russian recognition is the default; multilingual models also support English, Kazakh, Kyrgyz and Uzbek. Inference runs on your machine, with no cloud API, account or API key. Models are downloaded once and can be provisioned offline.
This documentation follows main. For an installed release, use the documentation at its tag. Model bundles and language-binding releases have their own versions.
Install
Choose one installation method:
- Prebuilt binary: download the archive for your platform from
Releases, unpack it and put
gigastt(gigastt.exeon Windows) onPATH. Builds cover macOS Apple Silicon, Linux x86_64/aarch64 and Windows x86_64. See release verification. - Homebrew:
brew tap ekhodzitsky/gigastt https://github.com/ekhodzitsky/gigastt brew install gigastt - Build the published crate:
Requires Rust 1.94+, a native toolchain andcargo install gigastt --lockedprotoc; Linux also needs OpenSSL development files. Build prerequisites.
Docker and offline/systemd packages: deployment.
Transcribe a file
gigastt transcribe recording.wav
gigastt transcribe recording.wav --format srt --output recording.srt
The first run downloads missing models. gigastt download preloads the ASR
bundle and, in default builds, the speaker model; punctuation is fetched when
enabled by serve or transcribe. See model setup and sizes.
Batch processing and directory watching:
gigastt transcribe-batch recordings/ transcripts/
gigastt watch inbox/ transcripts/
Files include WAV, AAC/M4A, MP3, FLAC, OGG/Vorbis, OGG/Opus and WebM/Opus. WAV telephony codecs and raw G.711/G.722 are supported; see the format and codec reference.
Run the server
gigastt serve
From another terminal, check readiness and upload a file:
curl --fail http://127.0.0.1:9876/ready
curl --fail http://127.0.0.1:9876/v1/transcribe \
-H 'Content-Type: application/octet-stream' \
--data-binary @recording.wav
Wait for /ready to return HTTP 200 before sending audio. /health also
answers during model loading. The default bind is loopback; remote access needs
explicit configuration and a proxy for TLS/authentication. See deployment.
| Interface | Purpose |
|---|---|
| CLI | Files, batch and watch; text, JSON, SRT, VTT and Markdown exports |
POST /v1/transcribe |
File upload with word timings and optional speaker labels |
POST /v1/transcribe/stream |
File upload with SSE results |
GET /v1/ws |
Live mono PCM16 at 8, 16, 24, 44.1 or 48 kHz; partial/final transcripts |
POST /v1/audio/transcriptions |
OpenAI-compatible multipart upload and output formats |
/v1/jobs |
Optional in-memory job queue, progress and cancellation (--enable-jobs) |
API reference · CLI reference · Go client · TypeScript client
Choose a model
| Model | Languages | Text output |
|---|---|---|
rnnt (default) |
Russian | Bare recognition; CLI/server can add punctuation and number normalization |
e2e_rnnt |
Russian | Built-in punctuation, casing and text normalization |
ml_ctc, ml_ctc_large |
Russian, English, Kazakh, Kyrgyz, Uzbek | Bare multilingual text |
Select with --model-variant. Models explains downloads,
formatting policy and upgrades; Languages reports measured
multilingual quality. Choose using recognition quality, speed and memory on your
workload, as well as punctuation.
Performance
The current main-branch RNN-T ASR bundle is about 310 MiB on disk. It uses INT8 matrix weights with floating-point convolutions; side models and optimized caches add space. Disk size is not a RAM requirement. The older v2.22.0 release pins a different bundle; see the model transition.
Benchmarks records corpus, model and hardware provenance. Historical M1 figures do not describe the new bundle. Live WebSocket recognition uses an offline model with buffered incremental decoding, so file WER and speed must not be presented as live-stream measurements.
Embed the engine
Use gigastt-core = "2.22" in Rust, npm install gigastt for the Node native
binding, or pip install gigastt for Python. These packages load side-provisioned
models. Swift, Kotlin and C ABI options are covered in
in-process quickstarts and packaging.
Package availability and versions differ across ecosystems.
CPU is the default backend. CUDA, CoreML, NNAPI and experimental ANE/Candle builds are described in architecture.
Documentation and development
Documentation index · Troubleshooting · Contributing · Changelog · Privacy · Security
GitHub workflows run only on manual dispatch. Local validation and the manual review/release process are documented in Contributing.
MIT. ASR weights are MIT; the optional WeSpeaker model is CC BY 4.0. See LICENSE, NOTICE and benchmark data licenses.