gigastt-uniffi

Idiomatic Swift, Kotlin, and Python bindings for gigastt — on-device Russian speech-to-text — generated from one Rust source with UniFFI.

Wraps the synchronous gigastt-core engine: models are side-loaded (no HTTP download dependency) and inference uses the blocking pool path (no tokio runtime), so the bindings are lean. Errors are typed (GigasttError → Swift throws / Kotlin exceptions / Python exceptions) and objects are reference-counted (no manual free).

API

process_chunk takes little-endian mono PCM16 and resamples to 16 kHz internally.

Generating the bindings

Build the cdylib, then run the version-pinned generator against it:

cargo build -p gigastt-uniffi
LIB=target/debug/libgigastt_uniffi.dylib   # .so on Linux

cargo run -p gigastt-uniffi --bin uniffi-bindgen -- generate --library "$LIB" --language python --out-dir bindings/python
cargo run -p gigastt-uniffi --bin uniffi-bindgen -- generate --library "$LIB" --language swift  --out-dir bindings/swift
cargo run -p gigastt-uniffi --bin uniffi-bindgen -- generate --library "$LIB" --language kotlin --out-dir bindings/kotlin

Generated bindings are build artifacts (bindings/ is git-ignored). Packaging them into a SwiftPM .xcframework, an Android .aar, and a PyPI wheel is the next step (prebuilt-artifacts task).

Python (quickstart, verified)

Install the prebuilt wheel — pip install gigastt — no compiler, no protoc, no onnxruntime download (it is statically linked; the wheel is py3-none-<platform>, one per platform across all Python 3.x). The ~215 MB model is side-loaded at runtime. Wheels are built + published by .github/workflows/python-wheels.yml.

# pip install gigastt
import gigastt_uniffi as g

engine = g.Engine("/path/to/gigastt/models")        # side-loaded model dir
t = engine.transcribe_file("recording.wav")
print(t.text)                                       # -> "шестьдесят тысяч тенге сколько будет стоить"
for w in t.words:
    print(w.text, w.start_s, w.end_s, w.confidence)

# streaming
s = g.Stream(engine)
for seg in s.process_chunk(pcm16_bytes, 16000):
    print(seg.text)
print([seg.text for seg in s.flush()])

Errors surface as exceptions:

try:
    g.Engine("/no/such/dir")
except g.GigasttError.ModelNotFound as e:
    ...

Features

coreml / cuda / nnapi forward to gigastt-core for hardware acceleration. The core dependency is lean (file-decode only — no net, no async-pool).

License

MIT.