Small, fast, open-source audio models — speech-to-text with timestamped segments and speaker diarization for meetings, text-to-speech that exports straight to Telegram-ready OGG/Opus voice notes, voice activity detection, and language ID. CLI tools plus an OpenClaw skill for LLM agents. CoreML on Apple Silicon, ONNX on Linux + Windows. One Rust binary. No Python. No ffmpeg.
$ bun add -g @drakulavich/kesha-voice-kit
Capabilities
Four pipelines, one binary. Pick what you need; the rest stays out of your way.
25 languages via NVIDIA Parakeet TDT. ~19× faster than Whisper on Apple Silicon, ~2.5× on CPU. Auto language detection, optional VAD for long audio, timestamped segments, plain text or LLM-friendly TOON output.
Drop a meeting recording in, get back timestamped segments tagged with a stable
speaker ID. Powered by FluidAudio’s Sortformer model running on the Apple Neural
Engine via a Swift sidecar — ~4.6× real-time on a 71-minute Zoom call. v1.19
gives the cold ANE warmup extra headroom so the first --speakers call
on a fresh process no longer hits a false timeout.
Kokoro-82M for English (CoreML on Apple Silicon, ONNX elsewhere) with acronym
auto-expand (FBI, EPAM, JSON), Vosk-TTS for Russian with stress markers, plus ~180
macOS system voices. Auto-routes by detected language. SSML
<prosody rate>, <break>, and
<emphasis> work on every backend — v1.21 closed the Apple
Silicon gap via FluidAudio 0.14.7’s model-native speed control.
Synthesize straight into the format Telegram, WhatsApp, Signal and Discord expect for inline voice messages — ~24× smaller than WAV. Pick a bitrate; the engine handles the OggOpus container, no ffmpeg required.
--json --timestamps returns segments with start and
end. Plays nicely with subtitles, meeting tools, and chunked LLM
summarisation. Programmatic API: transcribeWithTimestamps().
107 languages from audio via SpeechBrain ECAPA-TDNN. Text language detection runs through a precompiled Swift sidecar on macOS — 30–50 ms cold or warm.
Silero VAD v5 trims silence on long, sparse recordings. Auto-enabled past 120 seconds of audio.
--stdin-loopOne TTS process, many requests. Models stay in memory across calls so each synthesis hits ~10 ms cold-start instead of ~800 ms. Ideal for chatbots, live agents, and long agentic loops.
kesha record
Capture voice notes from the default microphone and pipe straight into
kesha --json to close the transcribe-then-reply loop. No
ffmpeg, no separate recorder process.
Opt-in local-only command events with paths and user-supplied text redacted.
kesha doctor surfaces status; kesha logs manages them;
kesha support-bundle bundles them for support. Mode
retain-on-failure keeps logs only when a command actually fails.
English + Spanish, French, Italian, and Portuguese on every platform via
CharsiuG2P (ByT5-tiny ONNX, CC-BY 4.0) with per-language number and acronym
normalizers — OTAN, OVNI, FIFA read as words. Mandarin Chinese
runs natively on Apple Silicon via FluidAudio 0.14.8’s tone-aware ANE variant.
Russian stays on Vosk-TTS. --lang es-ES for Castilian Spanish.
Every failure path prints a stable, machine-readable error [CODE]: …
line. kesha-engine --error-codes-json dumps the whole taxonomy; a
docs drift gate keeps docs/errors.md honest. Agents and CI can branch
on why something failed instead of regexing log soup.
A single Rust binary for macOS arm64, Linux x64, and Windows x64. FluidAudio
(CoreML) on Apple Silicon via dedicated sidecars (diarization, TTS, language
detection); ort (ONNX) elsewhere. Symphonia decodes WAV, MP3, OGG/Opus, FLAC,
AAC, M4A; pure-Rust opus + flacenc handle encode
— royalty-free codecs, no C dep, no patent exposure. Linux .deb and
.rpm packages signed with Sigstore.
No ffmpeg. No Python. No native Node addons.
Plug straight into your LLM agent as a voice-processing skill. Give Claude (or any agent) ears and a voice in one install.
Hear it
Real output from kesha say — Kokoro-82M for English, Vosk-TTS for Russian.
SSML <prosody rate> renders the same text at different speeds without
resampling artifacts. v1.21 brought prosody to Apple Silicon Kokoro too — EN, RU,
and the v1.22 multilingual voices all honor <prosody>,
<break>, and <emphasis>.
“Kesha turns voice into text in a single command — no Python, no ffmpeg.”
kesha say --voice en-am_michael "…" --out demo.flac
Same text, x-fast prosody. Tempo only — pitch and timbre unchanged.
kesha say --voice en-am_michael --ssml '<speak><prosody rate="x-fast">…'
«Кеша превращает голос в текст одной командой — без зависимостей.»
kesha say --voice ru-vosk-m02 "…" --out demo.flac
Same text, slow prosody. Useful for accessibility and dictation.
kesha say --voice ru-vosk-m02 --ssml '<speak><prosody rate="slow">…'
Performance
Compared against Whisper large-v3-turbo across Russian and English
clips — every engine left to auto-detect language. CoreML on M-series lands at roughly
19× real-time; the ONNX path on x86 still beats Whisper by ~2.5×.
Quick start
Requires Bun ≥ 1.3 on macOS arm64 or Linux x64.
Skip if you already have it.
$ curl -fsSL https://bun.sh/install | bash
Pulls the engine and base STT models, then warms the Neural Engine on macOS
(--tts pulls Kokoro + multilingual voices, ~990 MB on darwin-arm64).
First transcribe runs in ~500 ms instead of a 25-second hang.
kesha init is the guided entry point for humans;
kesha install is the explicit one for CI and agents — same code
underneath.
$ bun add -g @drakulavich/kesha-voice-kit
$ kesha init # humans (interactive)
$ kesha install --tts # + multilingual voicesPipe-friendly. Stdout = transcript, stderr = errors.
$ kesha audio.ogg
Свободу попугаям! Свободу!# Plain transcript
$ kesha audio.ogg
# Text + language/confidence
$ kesha --format transcript audio.ogg
# Full JSON with lang fields
$ kesha --format json audio.ogg
# Compact, LLM-friendly TOON
$ kesha --toon audio.ogg
# Long / silence-heavy audio (auto past 120s)
$ kesha --vad lecture.m4a
# Timestamped segments (start/end per phrase)
$ kesha --json --timestamps interview.ogg
# Meeting transcript with speaker IDs (darwin-arm64)
$ kesha --json --vad --speakers meeting.m4a
# Warn if detected language differs
$ kesha --lang en interview.wav
# Record from default mic, then transcribe straight to JSON
$ kesha record | kesha --json# Opt-in TTS pack: Kokoro (EN) + Vosk-TTS (RU), ~990MB
$ kesha install --tts
# Speak — voice auto-picks from text language
$ kesha say "Hello, world" > hello.wav
$ kesha say "Привет, мир" > privet.wav
# OGG/Opus voice notes (drop straight into Telegram)
$ kesha say --format ogg-opus "Уже бегу!" --out voice.ogg
# Variable-speed playback via SSML <prosody rate>
$ kesha say --voice en-am_michael --ssml \
'<speak><prosody rate="x-fast">Read this fast.</prosody></speak>' \
--out fast.wav
# Use a macOS system voice
$ kesha say --voice Milena "Чао!"Want to hear what these voices sound like? Play the four samples ↗
import { transcribe, downloadModel } from "@drakulavich/kesha-voice-kit/core";
// One-time install
await downloadModel();
// Transcribe a file
const text = await transcribe("audio.ogg");
// With detection metadata
const result = await transcribe("audio.ogg", { format: "json" });
console.log(result.language, result.confidence);# Multiple files — headers per file, like `head`
$ kesha freedom.ogg tahiti.ogg
=== freedom.ogg ===
Свободу попугаям! Свободу!
=== tahiti.ogg ===
Таити, Таити! Не были мы ни в какой Таити! Нас и тут неплохо кормят.
# Pipe transcripts into your favorite LLM
$ kesha --toon meeting.m4a | llm "Summarize action items"
# Status of installed engine + models
$ kesha status
# Diagnose env, log status, and CoreML caches
$ kesha doctor
# Bundle logs + manifests for a bug report
$ kesha support-bundleWhat's inside
Each model is best-in-class for its task and runs through
kesha-engine — Rust, with CoreML sidecars on Apple Silicon and ONNX
everywhere else. Nine multilingual voices ship in one bundle as of v1.22.
Audio decoding via Symphonia — WAV, MP3, OGG/Opus, FLAC, AAC, M4A.
Languages
Need detection only? VoxLingua107 covers 107 languages.
Integrations
Give your LLM agent ears. Voice-processing skill, pre-wired.
Setup guide →
macOS launcher commands: transcribe selected audio, speak the clipboard. From the
raycast/ directory.
Import transcribe() and downloadModel() directly into any
Bun/Node project.
Behind a corporate proxy or fully offline? Point Kesha at your model mirror.
Mirror docs →Open source, MIT licensed, and improving every week. Star the repo or jump straight to the install.