Speech-to-text with word-level timestamps, text-to-speech, live dictation and language ID. Your audio stays on your machine, and models download only when you ask for them. CoreML/ANE on Apple Silicon; ONNX on Linux and Windows. No Python. No ffmpeg. No surprise downloads.
$ bun add -g @drakulavich/kesha-voice-kit
Capabilities
Transcription works after one install. Speech synthesis, VAD and speaker labels are separate installs you add when you need them.
25 languages via NVIDIA Parakeet TDT. ~19× faster than Whisper on Apple Silicon, ~10× on an x86 CPU. Auto language detection, optional VAD for long audio, word-level timestamps, 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 on the Apple Neural Engine,
running inside the engine process. kesha install --diarize installs
it with VAD and warms the compiled model, so the first --speakers
run skips the cold compile.
kesha mcp runs a local Model Context Protocol server over stdio with
four tools: transcribe_audio, synthesize_speech,
list_voices and list_languages. Add it to Claude Code
with claude mcp add kesha -- kesha mcp, or to Claude Desktop, Codex,
Gemini CLI or Cursor with one config entry. A missing engine or model fails the
tool with a kesha install hint; nothing downloads on its own.
Run kesha audio.ogg before installing anything and, in a terminal,
Kesha shows the install plan, asks once, and on yes installs and transcribes in
the same run. Piped and non-interactive runs keep the error and its install
hint.
Kokoro runs through CoreML/ANE on Apple Silicon and ONNX on Linux and Windows;
Vosk-TTS covers Russian, and macOS system voices need no model download. macOS
auto-routes from the text language. On Linux and Windows mostly-Cyrillic text
still goes to ru-vosk-m02; other languages use --lang
or --voice. Kokoro (ANE or ONNX) and Vosk support SSML
<prosody rate> and <break>;
<emphasis> is Vosk-only. macOS system voices reject
--ssml.
Synthesize straight into a Telegram-ready voice note, ~24× smaller than WAV.
The default is 32 kbps; --bitrate takes 6000 to 510000 bits/s and
--sample-rate takes 8000, 12000, 16000, 24000 or 48000. The engine
handles the OggOpus container, no ffmpeg required.
--json --timestamps adds a words array to each segment,
with times counted from the start of the file. Each word's end is predicted on its
own, so neighbouring words can overlap.
107 languages from audio via SpeechBrain ECAPA-TDNN. JSON and TOON also expose
textLanguage separately from spoken audioLanguage:
Apple’s recognizer on macOS and bundled tinyld elsewhere. Their confidence scales
are not comparable.
Silero VAD v5 trims silence on long, sparse recordings. Install it explicitly
with kesha install --vad; Kesha uses it automatically past 120
seconds once it is available.
On darwin-arm64, kesha record --live prints what you say to stdout.
With VAD installed, --auto-stop ends the recording after a second of
silence. Recording stops at 120 seconds unless you pass
--max-seconds. Ctrl-C still prints the transcript, and the audio is
kept as a WAV when Ctrl-C or SIGTERM ends the session or the transcript cannot be
delivered.
Local-only command events with paths and user-supplied text redacted. The
default mode, retain-on-failure, writes a log only when a command
fails, so passing commands leave no file. kesha doctor surfaces
status; kesha logs manages them; kesha support-bundle
bundles them for support.
Kokoro speaks English, Spanish, French, Italian and Portuguese; Vosk-TTS speaks
Russian. Chinese, Japanese and Hindi are darwin-arm64 only:
Chinese accepts Han text, and the Kokoro Hindi and Japanese voices take romanized
input. Devanagari or kana/kanji text goes to an installed macOS system voice
instead. Text in a script the voice cannot speak is refused with
E_SCRIPT_UNSUPPORTED and a hint naming the script to transliterate
into. On Linux and Windows, pass --lang.
Set KESHA_HOME and the model cache, diagnostic logs, the stats
database, and MCP audio all live under one directory with the same layout on
every platform: cache/, logs/,
stats.sqlite, mcp-audio/. A per-location variable such
as KESHA_CACHE_DIR still wins; unset, nothing changes and nothing is
migrated. kesha status --json and kesha doctor --json
report each path with its source. On darwin-arm64 some FluidAudio files still
stay outside KESHA_HOME.
Every failure path prints a stable, machine-readable error [CODE]: …
line, and every engine progress, warning, or debug line on stderr is one NDJSON
event the CLI renders for you. kesha-engine describe publishes the
whole contract (flags, feature gates, the error taxonomy), and the CLI checks
each argv against it before spawning, so an unknown flag is refused with a
did-you-mean and exit 2. A run stopped by Ctrl-C or a signal reports
E_INTERRUPTED. Agents and CI can branch on why something
failed instead of regexing log soup.
kesha status --check
Readiness by exit code, for scripts. It exits 0 with no output when the engine
and the ASR model are installed. Otherwise it exits 1 with
E_MODEL_MISSING or E_ENGINE_SPAWN, or with
E_ENGINE_PROTOCOL when the engine does not answer within 15 seconds.
Plain status and status --json still exit 0.
A single Rust binary for macOS arm64, Linux x64, and Windows x64. FluidAudio
(CoreML) runs in-process on Apple Silicon; two Swift sidecars cover macOS system
voices (say-avspeech) and text-language detection
(kesha-textlang); ort (ONNX) elsewhere. Kesha reads ogg, opus, mp3,
m4a, aac, wav, flac, aiff and webm. Encoding uses one statically linked libopus
and pure-Rust FLAC: royalty-free codecs, no patent exposure. Linux .deb and
.rpm packages signed with Sigstore.
No ffmpeg. No Python. No native Node addons.
Kesha ships as an OpenClaw plugin. Point OpenClaw's audio tool at
kesha and voice messages reach your agent as text, transcribed
locally with no API keys.
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. Kokoro and Vosk honor <prosody> and
<break>; <emphasis> is Vosk-only.
“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.
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, with every engine left to auto-detect the language. CoreML on M-series is roughly
19× faster than openai-whisper; the ONNX path on an x86 CPU is ~10× faster than
openai-whisper and ~2.5× faster than faster-whisper.
Quick start
Requires Bun ≥ 1.4.4 on macOS arm64, Linux x64, or Windows x64. Engine and models download only when you ask them to.
Skip if you already have it.
macOS / Linux
$ curl -fsSL https://bun.sh/install | bashWindows PowerShell
> powershell -c "irm bun.sh/install.ps1 | iex"
Preview the size, then explicitly download the engine and base STT models: about
0.6 GB on Apple Silicon, 2.5 GB on Linux and Windows. kesha init is
the guided alternative. Progress and recovery hints go to stderr. Skip this step
and the first kesha audio.ogg in a terminal shows the plan and asks
once before it downloads; piped runs keep the error. Add --tts or
--vad only when those workflows need them.
$ bun add -g @drakulavich/kesha-voice-kit
$ kesha install --plan && kesha install
$ kesha install --vad # only for live auto-stop / VADPipe-friendly: stdout is the transcript; stderr carries progress and 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 (after: kesha install --vad)
$ kesha --vad lecture.m4a
# Segments plus word-level timestamps
$ kesha --json --timestamps interview.ogg
# Meeting transcript with speaker IDs (darwin-arm64)
$ kesha install --diarize && kesha --json --speakers meeting.m4a
# Warn if detected language differs
$ kesha --lang en interview.wav
# Live dictation: darwin-arm64 + explicit VAD install
$ kesha install --vad && kesha record --live --auto-stop# Opt-in TTS pack: English; preview sizes first
$ kesha install --tts
# macOS auto-routes from text language
$ kesha say "Hello, world" > hello.wav
$ kesha say "Привет, мир" > privet.wav
# Linux / Windows: state the language explicitly
$ kesha say --lang ru "Привет, мир" > privet.wav
# OGG/Opus voice notes (drop straight into Telegram)
$ kesha say --format ogg-opus "Уже бегу!" --out voice.ogg
# Kokoro / Vosk only: 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 { install, transcribe, say, KeshaError } from "@drakulavich/kesha-voice-kit/core";
// The one call that downloads: engine, ASR models, English TTS
await install({ tts: ["en"] });
// Transcribe a file: resolves to a TranscribeResult
const { text, lang } = await transcribe("audio.ogg");
// Speak: resolves to the audio bytes
const wav = await say({ text: "Hello" });
// Every rejection is a KeshaError with a stable code
try {
await transcribe("note.ogg");
} catch (err) {
if (err instanceof KeshaError && err.code === "E_ENGINE_SPAWN") console.error(err.hint);
else throw err;
}# 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 runs through kesha-engine, written in Rust: CoreML
in-process on Apple Silicon and ONNX everywhere else. Text-to-speech covers nine
languages across Kokoro and Vosk.
Audio decoding via Symphonia, with a libopus adapter for Opus: ogg, opus, mp3, m4a, aac, wav, flac, aiff and webm.
Languages
Need detection only? VoxLingua107 covers 107 languages.
Integrations
Voice messages reach your OpenClaw agent as text, transcribed locally. Ships as a plugin.
Setup guide →Dictate to Clipboard on macOS: records with a live meter, auto-stops on silence, transcribes locally, and copies the text.
Raycast Store →
Import transcribe(), say(), install() and
capabilities() into a Bun program. transcribe()
resolves to a TranscribeResult, and every rejection is a
KeshaError with a stable code.
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.