otoji (音字)

realtime speech ⇄ text — 音を字に

otoji is a Rust workspace that wires up streaming ASR, LLM-polished transcripts, and TTS behind a single react-ink-style terminal UI built on ratatui.

mic / file ──► AudioChunk ──► AsrProvider ──► AsrEvent ──► Polisher ──► TUI
                                                                    └─► transcript.md

Workspace layout

Crate	Purpose
otoji-core	Shared types: AudioChunk, AsrEvent, Word, OtojiError
otoji-audio	Audio sources — cpal mic capture (with resampling) and PCM file replay
otoji-asr	AsrProvider trait + iflytek_rtasr (HMAC-SHA1 signa, WebSocket)
otoji-tts	TtsProvider trait + iflytek_tts (HMAC-SHA256 auth, MP3/PCM streaming)
otoji-polish	Polisher trait + NoopPolisher and AnthropicPolisher (Claude Haiku 4.5 default)
otoji-cli	otoji binary — clap subcommands + ratatui TUI

See ./docs/ for the architecture rationale and the comparison of RT ASR providers (iFlytek RTASR / CoLi / SenseVoice / Whisper / Deepgram).

Build

cargo build --release

Usage

# 1) Live mic → RTASR → polished TUI
export IFLYTEK_APP_ID=...
export IFLYTEK_API_KEY=...
export ANTHROPIC_API_KEY=...   # optional, enables LLM polish layer
cargo run -p otoji-cli -- listen

# 2) Replay a 16kHz mono PCM file in real time
cargo run -p otoji-cli -- file 16k_10.pcm

# 3) Synthesize speech via iFlytek TTS
export IFLYTEK_TTS_API_KEY=...
export IFLYTEK_TTS_API_SECRET=...
cargo run -p otoji-cli -- speak "你好，世界" --out hello.mp3

TUI

The transcript view shows:

• [seg_id] confirmed segments in white bold (polished) or gray (raw, awaiting polish)
• The current partial hypothesis as ░ ... in dark gray italic
• A header with provider state and counts

Press q / Esc / Ctrl-C to quit.

Roadmap

• otoji-asr/coli.rs — CoLi ASR via ListenHub
• otoji-asr/sensevoice.rs — FunASR self-host bridge
• otoji-tts/edge_tts.rs — Microsoft Edge TTS as a free fallback
• otoji-cli record — write transcripts to *.md next to the source audio
• Bench harness (CER / latency / cost) under crates/otoji-bench

License

MIT