Kiku (聞く, to listen) — Transcription Engine Built in Rust

what:

kiku is a minimal transcription engine written in rust. audio in, transcript out — that’s the whole contract. it powers the transcription layer of kioku, but it’s built to be generally usable by anyone.

why:

kioku’s original transcription-service was a python fastapi app wrapping faster-whisper, and over time it accumulated everything: admission control, concurrency semaphores, tier queues, retry logic, silence heuristics, three STT backends. the engine and the orchestration were living in the same file.

tl;dr it was really sloppy and i couldn’t comprehend any logic behind of the things, well i use AI too for dev, but i often use it in like mentor-mode to really let myself form a mental model first and learn about rust in general, before actually shipping it

so the split: kiku is the engine, kioku is the orchestrator. all the tier management, queueing, and scaling policy stays in kioku.

design:

one enum, two backends:

• cloud — OpenRouter’s /audio/transcriptions endpoint. one call, any STT model they route: chirp-3, gpt-4o-transcribe, whisper-large-v3.
• local — whisper.cpp through whisper-rs, behind a local-whisper cargo feature so the default build needs no C++ toolchain.

let engine = Engine::from_config(&cfg)?; // model loads once, here
let t = engine.transcribe(&audio, "wav", None).await?; // call forever

went with an enum instead of a trait object — async fn in traits isn’t dyn-safe without the async-trait crate, the backends are known at compile time, and the compiler yells if a variant forgets to implement. boring wins.

some small things i like:

• one MODEL env var for both worlds: an openrouter id (openai/gpt-4o-mini-transcribe) or a ggml name (large-v3, tiny) — local models auto-download from ggerganov/whisper.cpp on first run.
• any PCM wav works: multichannel gets downmixed, non-16kHz gets resampled with a hand-rolled linear interpolation resampler (~15 lines, plenty for speech).
• fine-tunes: a single shell script converts any HF transformers whisper fine-tune to ggml. i’m testing cobrayyxx/whisper-small-indo-eng for indonesian-english meetings.
• tests without the network: the integration tests spin up a fake OpenRouter with axum on port 0, capture the request body, and assert the base64 audio round-trips. no mock library, axum was already there.

status:

shipped on crates.io!

https://crates.io/crates/kiku

enjoy the cargo install kiku :)