kesha-voice-kit

Kesha Voice Kit

Give your local tools and LLM agents a voice.
Fast speech-to-text, text-to-speech, voice-activity detection, and language detection in one local-first CLI — CoreML on Apple Silicon, ONNX on Linux and Windows.

• Transcribe locally — 25 languages, up to ~19x faster than Whisper on Apple Silicon, ~2.5x on CPU
• Speak back — text-to-speech in 9 languages
• Plug into agents — ship voice workflows as CLI commands, an MCP server, an OpenClaw skill, or a Hermes agent
• Small Rust engine — single ~65MB binary, no ffmpeg, no Python, no native Node addons

Quick Start

Runtime: Bun >= 1.3.0.

# 1. Install Bun (skip if you have it)
curl -fsSL https://bun.sh/install | bash        # macOS/Linux — or: brew install oven-sh/bun/bun
powershell -c "irm bun.sh/install.ps1 | iex"    # Windows

# 2. Install Kesha
bun add -g @drakulavich/kesha-voice-kit
kesha --version                                 # confirms `kesha` resolved on PATH

# 3. Download the engine and models — pick one path
kesha init                                      # guided: TTS languages and optional VAD / diarization
kesha install --plan && kesha install           # manual: preview the sizes, then download

# 4. Transcribe
kesha audio.ogg                                 # transcript to stdout

kesha install pulls ~2.5 GB on Linux/Windows and ~0.6 GB on Apple Silicon, whose CoreML engine reads a smaller model set. It is always explicit — nothing downloads behind your back — and reports download progress on stderr. If bun --version fails right after step 1, reload your PATH: exec $SHELL -l.

Prefer Homebrew or Docker? See Other install methods. Air-gapped or behind a corporate mirror? See docs/model-mirror.md.

Platform support

All three targets transcribe, detect the spoken language, run VAD, and speak. The macOS-only rows need Apple frameworks — they are not a missing port. Windows is a tested path rather than a published binary nobody ran: CI does a cold kesha install on windows-latest, transcribes a fixture, and round-trips a synthesis (#216, #667).

	macOS arm64	Linux x64	Windows x64
Transcribe · audio language ID · VAD	CoreML / ANE	ONNX CPU	ONNX CPU
TTS — en ru es fr it pt	✅	✅	✅
TTS — hi ja zh and macOS system voices	✅	—	—
Mic capture and live dictation (kesha record)	✅	—	—
Speaker diarization (--speakers)	✅	—	—
Word-level timestamps (words in --json)	✅	✅	✅
Voice auto-routing from the text's language	✅	pass --lang	pass --lang

Intel Macs get no published engine binary. Full matrix with maturity labels: docs/product-positioning.md.

Speech-to-text

kesha audio.ogg                            # transcribe (plain text)
kesha --format transcript audio.ogg        # text + language/confidence
kesha --format json audio.ogg              # full JSON with lang fields
kesha --json --timestamps audio.ogg        # JSON with timestamped segments
kesha --itn audio.ogg                      # spelled-out numbers -> digits
kesha --toon audio.ogg                     # compact LLM-friendly TOON
kesha status                               # show installed backend info
kesha status --disk                        # + recursive cache disk usage
kesha status --json                        # machine-readable, for scripts

Multiple files get head-style headers; stdout is the transcript, stderr is errors — pipe-friendly:

$ 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.

• Record from the mic (macOS): kesha record --out hello.wav writes microphone audio to a WAV file (kesha hello.wav transcribes it). macOS prompts for microphone access on first use — grant it under System Settings → Privacy & Security → Microphone if it was denied. On Linux/Windows or headless boxes, pass any existing audio file straight to kesha instead.
• Dictate straight to text (darwin-arm64): kesha record --live transcribes the mic as it captures and prints the transcript to stdout — no WAV in between, so it pipes (kesha record --live | pbcopy). To end after trailing silence, explicitly install VAD then opt in: kesha install --vad && kesha record --live --auto-stop. The defaults are 1,000 ms of silence after 250 ms of speech; tune them with --auto-stop-silence-ms, --auto-stop-min-speech-ms, and --auto-stop-threshold. Progress goes to stderr. Linux and Windows do not capture the microphone; pass an existing audio file to kesha to transcribe it. An interruption is recoverable: Ctrl-C (or SIGTERM) stops the session, still prints what you dictated, and exits 130/143, and the audio is spilled to a recovery WAV under ~/.cache/kesha/recordings/ — named on stderr when the session starts, deleted once the transcript has actually been delivered, kept if anything — a signal, a crash, a closed terminal, a dead pipe — got in the way first (#962).
• Long / silence-heavy audio: install VAD (kesha install --vad); Kesha auto-uses it past 120 s. Without VAD, long audio falls back to fixed ASR chunks. See docs/vad.md.
• Speaker diarization (darwin-arm64): kesha install --diarize (which installs VAD too), then kesha --json --speakers meeting.m4a stamps each segment with a speaker id. --speakers engages VAD windowing itself at any duration, so it cannot be combined with --no-vad. Linux/Windows return a clear "darwin-arm64 only" error (#199).
• Word-level timestamps (every platform): kesha --json --timestamps audio.ogg adds a words array to each segment — { "word": "email", "start": 0.72, "end": 1.12 } — on the same file-relative clock as the segment, so a word always lies inside the segment carrying it. Read them off the decoder's own frame grid, so: times are quantised to 0.08 s, consecutive spans may overlap (each end is a per-word duration prediction, not the next word's start), end >= start rather than strictly greater, and punctuation stays attached to its word. The key is simply absent where a segment has none — any segment --itn rewrote, for one — so check the transcribe.words capability rather than expecting an empty array (#720).
• Text-language detection: JSON and TOON results include textLanguage with a language code, confidence, and its source. On macOS Kesha uses Apple NLLanguageRecognizer; elsewhere it uses the bundled tinyld fallback, whose confidence scale is different. This is separate from audioLanguage, which identifies the spoken audio when available.
• Written-form numbers: --itn rewrites what the model spells out — "two hundred thirty two" → "232", "five dollars and fifty cents" → "$5.50". Opt-in, every platform, timestamps untouched. English-only in practice; Russian and the rest pass through unchanged. Spoken punctuation names stay words ("dot", "comma", "the period of growth") because Kesha transcribes speech rather than dictation — so "example dot com" keeps its words too (#822). A sentence "and" survives the number that follows it ("cats and three dogs" → "cats and 3 dogs"), while an "and" the number owns still joins it ("three hundred and five" → "305") (#1000) — and no longer splits the number around it ("two hundred and thirty two" → "232", not "230 2") (#1006). A hyphenated number reads the same as the spaced form ("twenty-five apples" → "25 apples"), while a hyphen between ordinary words is left alone ("well-known", "state-of-the-art", "twenty-something") (#1004).

Text-to-speech

Kesha speaks back in 9 languages. Kokoro runs natively through FluidAudio CoreML/ANE on Apple Silicon and through ONNX on Linux and Windows; Russian uses Vosk-TTS, while macos-* system voices need no model download. On macOS Kesha picks the voice from the text's own language; on Linux and Windows, state the language with --lang <code> (or the voice with --voice <id>) — otherwise the engine default speaks.

kesha install --tts                              # English voices; sizes differ per platform — preview: kesha install --plan
kesha install --tts en ru                        # + Russian (+~890 MB, Vosk)
kesha say "Hello, world" > hello.wav
kesha say "Привет, мир" > privet.wav             # auto-routes by language (macOS)
kesha say --lang ru "Привет, мир" > privet.wav   # explicit — the Linux/Windows path
kesha say --voice ru-vosk-m02 "Голос в текст." > ru.wav

Output formats (--format, or inferred from the --out extension):

kesha say "Hello" --out hi.wav                    # WAV (default, uncompressed)
kesha say "Hello" --format ogg-opus --out hi.ogg  # OGG/Opus — messenger voice notes
kesha say "Hello" --format flac --out hi.flac     # FLAC — lossless, plays in every browser incl. Safari/iOS

kesha say --list-voices lists what's installed. Voices, the full catalogue, macOS system voices, SSML, speaking rate (--rate, <prosody>), Russian word stress, and Russian/English abbreviation handling are all in docs/tts.md.

Languages

Speech-to-text spans 25 languages and text-to-speech 9 — full tables with codes, flags, and per-platform availability in docs/languages.md. Audio language detection identifies 107 languages.

Performance

Up to ~19x faster than Whisper on Apple Silicon (M2), ~2.5x faster on CPU

Compared against Whisper large-v3-turbo, all engines auto-detecting language:

Full per-file breakdown (Russian + English): BENCHMARK.md. The CPU figure is the ONNX engine on an M2's CPU cores; no x86 numbers are published yet.

Other install methods

All of these install the Bun CLI wrapper; engine + models still download explicitly via kesha install. (Nix is the exception — it currently builds only the engine from source; see below.)

• Homebrew — brew install drakulavich/tap/kesha-voice-kit · docs/homebrew.md
• Linux packages (.deb/.rpm, x64) — published on CLI releases, see docs/linux-packages.md
• Docker (GHCR image) — docs/docker.md
• Nix (aarch64-darwin / x86_64-linux) — builds the engine from source (nix build github:drakulavich/kesha-voice-kit#kesha-engine). The full kesha CLI via nix run / nix profile install is not yet available — it needs a maintainer with Nix to populate a build hash (#946). · docs/nix-install.md
• Shell completions + manpage — kesha completions bash|zsh|fish and kesha manpage print the packaged files to install wherever your shell expects them.

Integrations

• MCP server — kesha mcp exposes transcribe/synthesize/list tools to any MCP client (Claude, Cursor, Codex, Gemini). Setup: docs/mcp.md.
• OpenClaw — give your LLM agent ears. Install & config: docs/openclaw.md.
• Hermes Agent — local STT/TTS through Hermes command providers. Setup: docs/hermes.md.
• Raycast (macOS) — offline microphone dictation from the launcher: Dictate to Clipboard records with a live signal meter, auto-stops on silence, transcribes locally, and copies the text. Install from the Raycast Store · source: raycast/.
• Programmatic API — @drakulavich/kesha-voice-kit/core for use inside a Bun program. See docs/api.md.

More

• Architecture — runtime data flow, the models that ship, the CLI ↔ Rust engine boundary, model pinning, and where tests live.
• Use cases — copy-paste recipes (transcribe a meeting, speak from OpenClaw, run offline, move the cache).
• Product positioning — supported workflows, non-goals, maturity labels, platform matrix.
• Changelog — every release, with the behaviour changes spelled out.
• Diagnostics: kesha doctor, kesha support-bundle (redacted .tar.gz for issues), and kesha logs produce local, content-free diagnostics — see docs/diagnostic-logs.md. Every failure prints a stable error [CODE]: … line and a documented process exit code.
• Scripting & CI: --json (or --toon) for machine-readable output, --include-errors (with either) to get per-file failures on stdout alongside the results, --quiet/-q to silence progress, and --no-color (or NO_COLOR=1) for plain logs. Colors switch off automatically when CI=true.
• Privacy / Local Stats: Stats are off by default and fully local. Opt in with kesha stats enable to record content-free operational metrics in a local SQLite database — never networked, never storing audio, transcripts, text, or paths. Full commands & lifecycle: docs/local-stats.md.

Contributing

See CONTRIBUTING.md, the Roadmap (Now / Next / Later), and the Decision log (why platform/model choices were made — and reversed). Dev setup: just dev-setup (Bun, Rust, nextest, platform libs).

License

Made with 💛🩵 and 🥤 energy under MIT License