Impossible Voice

Impossible Voice is a ready-made, self-hosted speech-to-text and text-to-speech server. The first release targets one curated local STT model and one curated local TTS voice with automatic, checksum-verified installation and offline operation after setup.

The public v0.1 promise is frozen in docs/product-contract.md. HTTP, realtime WebSocket, gRPC, and bounded MCP transports are implemented. Native Windows and Linux archives are built by the release workflow; a prebuilt Linux container is not published in v0.1.

From a source checkout, install the pinned local artifacts with one command, then start the server with one command.

PowerShell:

./scripts/setup.ps1
./scripts/serve.ps1

Bash:

./scripts/setup.sh
./scripts/serve.sh

The equivalent direct CLI is:

cargo run --locked --bin impossible-voice -- setup
cargo run --locked --bin impossible-voice -- status
cargo run --locked --bin impossible-voice -- doctor
cargo run --locked --bin impossible-voice -- serve

setup installs exactly one pinned sherpa-onnx CPU runtime, one English streaming STT model, and one English TTS voice into the ignored runtime-artifacts directory. Downloads are HTTPS-only, size- and SHA-256-verified, safely extracted into staging, fully inventoried, then atomically activated. setup --offline re-verifies an existing installation without network access; status is read-only. Ordinary serve never downloads artifacts.

The release workflow publishes a Windows ZIP and a mode-preserving Linux tar.gz, each with a SHA-256 sidecar. They contain the native server, scripts, configuration, documentation, and notices—but no runtime or model bytes. See docs/operations.md for archive use, checksum verification, offline restart, and recovery.

serve binds HTTP/WebSocket/MCP to 127.0.0.1:8080 and gRPC to 127.0.0.1:50051 by default. Configuration precedence is defaults, then an optional TOML file, then IMPOSSIBLE_VOICE_* environment variables, then explicit CLI flags. See config/impossible-voice.example.toml. Readiness remains false until the curated artifacts and both engines are usable.

Batch transcription:

curl -F "file=@sample.wav;type=audio/wav" http://127.0.0.1:8080/v1/audio/transcriptions

Speech synthesis:

curl -H "content-type: application/json" \
  -d '{"input":"Impossible Voice is local.","voice":"kristin","response_format":"wav"}' \
  http://127.0.0.1:8080/v1/audio/speech --output speech.wav

The exact HTTP, WebSocket, gRPC, and MCP contracts and bounded examples are documented in docs/api.md. WebSocket and gRPC TTS chunks are emitted after native synthesis finishes; they are bounded transport chunks, not incremental model generation.

Development

The repository is a Rust workspace with isolated audio, artifact, native FFI, STT, TTS, protocol, and server crates.

cargo fmt --all -- --check
cargo check --locked --workspace --all-targets
cargo clippy --locked --workspace --all-targets -- -D warnings
cargo test --locked --workspace
pwsh ./scripts/test-privacy-scan.ps1
pwsh ./scripts/privacy-scan.ps1
pwsh ./scripts/verify-vendored-impossible-server.ps1
pwsh ./scripts/test-vendored-impossible-server.ps1
cargo test --locked --manifest-path vendor/impossible-server/Cargo.toml --target-dir target/foundation

The vendored foundation is a deliberately narrow, integrity-checked snapshot of the reusable core and testkit. It supplies lifecycle and safety primitives; it does not supply Voice transports, WebSockets, audio processing, inference engines, or artifact installation. See docs/foundation-vendoring.md.

License

Licensed under either the Apache License, Version 2.0 or the MIT license, at your option.