kaiku 1.2.2

Local-first speech-to-text CLI: capture, denoise, transcribe, post-process, clipboard.

kaiku -- Modular Voice (SST) Pipeline Prototype :)

Record speech, transcribe it, and copy the result to clipboard or a file. Supports cloud and fully local ASR backends, VAD (for background daemons) and toggle (for keyboard shortcuts), noise reduction, speaker diarization (as an ASR backend), and AI post-processing.

Originally forked from the elegant but minimalist speech to clipboard tool asr2clip which is also available from . The word "kaiku" is Finnish for "echo".

Jump to the Related projects section at the end to understand the landscape of ASR related tooling and why this project was developed for the open source community.

TL;DR

Try mock backends first (no API keys, no whisper.cpp build):

pip3 install kaiku
kaiku --download-fixtures # ~1.3 MB: demo WAVs + transcripts; prints config paths
kaiku --generate-config   # create config with all backend examples
# paste fixture paths from download output into mock_devices / mock-fwd in config
kaiku --test
kaiku -d mock-jfk -b mock-fwd                 # both mock audio and mock transcription

Cloud (API) path:

pip3 install kaiku
kaiku --generate-config   # create config with all backend examples
kaiku --edit              # fill in your API key
kaiku --test              # verify
kaiku                     # record and transcribe

Local offline path — sherpa-onnx with VAD support (model auto-downloads):

pip3 install kaiku[vad]
kaiku --download-model    # download SenseVoice model on first use
kaiku --serve &           # start local ASR API server
# configure a backend pointing to http://127.0.0.1:8000/v1/ — see Local ASR server below
kaiku --test -b sonnx
kaiku -b sonnx

Local offline path — whisper.cpp (no VAD):

pip3 install kaiku
# build whisper.cpp and download the models you want, then configure it in config
kaiku --generate-config   # shows a wcpp backend example
kaiku --test -b wcpp
kaiku -b wcpp

CLI reference

usage: kaiku [-h] [-v] [-q] [-c FILE] [-e] [--generate-config]
                [--print-config] [--test] [-x NAME] [--list-devices] [-d DEV]
                [-i FILE] [-p NAME] [-b NAME] [-l LANG] [-r] [-C SEC] [-g]
                [--serve] [--host HOST] [--port PORT] [--model-dir MODEL_DIR]
                [--num-threads NUM_THREADS] [--download-model] [--vad]
                [--interval SEC] [--silence-threshold PROB]
                [--silence-duration SEC] [-s N] [-P NAME] [-M MODEL] [-o FILE]
                [-T NAME] [-z]

Record audio and transcribe to clipboard using ASR API

optional arguments:
  -h, --help            show this help message and exit
  -v, --version         show program's version number and exit
  -q, --quiet           Quiet mode — only output transcription and errors

Setup:
  -c FILE, --config FILE
                        Path to configuration file
  -e, --edit            Open configuration file in editor (creates default
                        config if missing)
  --generate-config     Write config template to
                        ~/.config/kaiku/config.yaml
  --print-config        Print config template to stdout
  --test                Test backend connectivity and configured
                        preprocessors, then exit
  -x NAME, --preset NAME
                        Pipeline preset name (key under 'presets:' in config).
                        Presets define complete pipelines: ASR backend,
                        preprocessor, post-processor. Optional if
                        'default_preset' is set in config; CLI overrides still
                        work (-b, -p, -P).

Audio:
  --list-devices        List available audio input devices
  -d DEV, --device DEV  Audio input device (name, ALSA name, or index).
                        Overrides config.
  -i FILE, --input FILE
                        Transcribe an existing audio or video file instead of
                        recording. Supported: wav, mp3, m4a, ogg, flac, aac,
                        opus, wma, mp4, mov, mkv, webm, avi, flv, mvi
  -p NAME, --preprocessor NAME
                        Audio preprocessor: none, noisereduce, pyrnnoise,
                        deepfilter. Overrides the preprocessor in the selected
                        preset.

Transcription:
  -b NAME, --backend NAME
                        ASR backend to use (key under 'asr_backends:' in
                        config). Overrides the backend in the selected preset.
  -l LANG, --language LANG
                        Language hint for transcription (ISO-639-1, e.g. 'fi',
                        'en'). Overrides config. Omit to auto-detect.
  -r, --robust          Robust mode for -i file input: split at silence
                        boundaries, quality-check chunks, retry failures,
                        stream output (tail-f friendly).
  -C SEC, --chunk-duration SEC
                        Max chunk duration in seconds for -r/--robust mode
                        (default: 180)
  -g, --toggle          Toggle recording: first call starts, second call stops
                        and transcribes. Designed for keyboard shortcuts.

Local ASR server:
  --serve               Start the local sherpa-onnx ASR API server
  --host HOST           Server bind address (default: 127.0.0.1 or
                        local_asr.host in config)
  --port PORT           Server bind port (default: 8000)
  --model-dir MODEL_DIR
                        Path to ASR model directory
  --num-threads NUM_THREADS
                        Inference threads (default: 4)
  --download-model      Download the SenseVoice model and exit

VAD (continuous recording):
  --vad                 Continuous recording with voice activity detection.
                        Transcribes automatically when silence is detected
                        after speech. Requires sherpa-onnx: pip3 install
                        kaiku[vad].
  --interval SEC        Continuous recording with fixed interval (seconds)
  --silence-threshold PROB
                        VAD speech probability threshold, 0.0-1.0 (default:
                        0.5)
  --silence-duration SEC
                        Silence duration to trigger transcription (default:
                        1.5 s)

Diarization:
  -s N, --speakers N    Speaker count hint for diarization backends (type:
                        whisperx). Ignored by all other
                        backends. Selects a diarization backend in your preset
                        or with -b / --backend; see 'asr_backends:' in config.
                        If omitted, the backend uses its own default or auto-
                        detects speaker count.

Post-processing:
  -P NAME, --post NAME  AI post-processor name (key in 'postprocessors:'
                        config) or an inline system-prompt string. Requires
                        'postprocessor_backends:' in config. Overrides the
                        post-processor in the selected preset.
  -M MODEL, --post-model MODEL
                        AI model used for the post-processing (f. ex. claude-
                        sonnet-4-6). Overrides the post-processor config for
                        this run.

Output:
  -o FILE, --output FILE
                        Append transcripts to file
  -T NAME, --template NAME
                        Output template name from 'output_templates:' in
                        config. Controls what is written to clipboard/-o FILE.
                        Overrides the template specified in the prompt
                        definition.
  -z, --no-clipboard    Do not copy the transcript (or a file path) to the
                        system clipboard. Stdout and -o output behave as
                        usual.

Examples:
  kaiku --edit                             # create/open config in editor
  kaiku --test                             # verify backend and preprocessors
  kaiku                                    # record, transcribe, copy to clipboard
  kaiku --toggle                           # toggle recording (for keyboard shortcuts)
  kaiku --toggle -P solo-restructure       # toggle, and produce AI-structured memo
  kaiku -i audio.mp3                       # transcribe an existing file
  kaiku -i m.mp3 -p deepfilter -r          # neural denoising + chunked transcription
  kaiku -i meeting.m4a -b whisperx -s 3    # speaker diarization, 3-speaker hint
  kaiku --serve                            # start local sherpa-onnx ASR server
  kaiku --vad -o meeting.txt               # continuous VAD transcription to file
  kaiku --interval 60                      # fixed-interval continuous recording

See https://github.com/sjjsy/kaiku for full documentation and configuration examples.

Prerequisites

Python 3.8+ and one of:

• Cloud API key (OpenAI Whisper, Groq, SiliconFlow, xinference, or any OpenAI-compatible endpoint)
• Local sherpa-onnx server (pip3 install kaiku[vad], model auto-downloads)
• Local whisper.cpp binary + model file (fully offline, no key needed)

System packages

Dependency	Purpose	Linux	macOS	Windows
ffmpeg	Audio format conversion	apt install ffmpeg	brew install ffmpeg	Download
PortAudio	Audio recording	apt install libportaudio2	brew install portaudio	Included with sounddevice

Optional Python extras

Extra	Install	Purpose
vad	pip3 install kaiku[vad]	VAD continuous recording + local sherpa-onnx ASR server
deepfilter	pip3 install kaiku[deepfilter]	DeepFilterNet3 best-quality noise reduction
noisereduce	pip3 install kaiku[noisereduce]	Spectral noise reduction (scipy)
pyrnnoise	pip3 install kaiku[pyrnnoise]	RNNoise GRU noise reduction (scipy)
enhance	pip3 install kaiku[enhance]	All three noise reduction options
diarize	pip3 install kaiku[diarize]	Speaker diarization via WhisperX

Installation

pip3 install kaiku

# or in an isolated environment
pipx install kaiku

# upgrade
pip3 install --upgrade kaiku

All extras: Noise reduction options + VAD and the local sherpa-onnx ASR server:

pip3 install kaiku[enhance,vad]

Note: Audio preprocessing (-p) is not (yet) applied in VAD/interval continuous mode.

From source

git clone https://github.com/Oaklight/kaiku.git
cd kaiku
pip3 install -e .

Setup

Setup commands manage your configuration file and verify that configured backends and preprocessors are working.

Flag	Description
-c FILE	Path to a specific configuration file
-e / --edit	Open config in editor (creates default if missing)
--generate-config	Write the annotated config template to ~/.config/kaiku/config.yaml
--print-config	Print the config template to stdout
--test	Test backend connectivity and preprocessor availability, then exit
-x NAME / --preset NAME	Pipeline preset (key under presets:). Optional if default_preset is set in config.
-q / --quiet	Suppress informational output; only print the transcript and errors

Setup commands

kaiku --generate-config   # write a fully annotated config with all backend examples
kaiku --edit              # create/open config in your default editor
kaiku --print-config      # print the annotated template to stdout
kaiku --test              # verify backend connectivity and preprocessors
kaiku --test -b wcpp      # test a specific backend

Config file

Config file is created at ~/.config/kaiku/config.yaml. Locations searched in order:

1. ./kaiku.conf
2. ~/.config/kaiku/config.yaml
3. ~/.config/kaiku.conf
4. ~/.kaiku.conf

Note: The created config file embeds partially commented-out configuration options along with brief explanations for most features. See kaiku/kaiku.conf.example in the repo for a complete, current example with all backend and feature documentation. The config file template is not usable immediately: You must update it based on your setup and needs. The following sections tackle some of these in more detail where relevant.

Audio

Flag	Description
--list-devices	List available audio input devices with names and indices
-d DEV / --device DEV	Audio input device (name, ALSA name, or index). Overrides audio_device in config.
-p NAME / --preprocessor NAME	Audio preprocessor: none, noisereduce, pyrnnoise, deepfilter. Overrides config.

Audio device

kaiku --list-devices            # list available input devices with names and indices
kaiku -d "plughw:Snowball"      # use a specific ALSA device for this run

System audio routing (recommended)

On Linux, set audio_device to "pulse" or "pipewire" and select the active microphone in pavucontrol or your desktop sound settings.

audio_device: "pulse"              # PulseAudio routes to whichever mic is set as default input

Targeting a specific device directly

audio_device: "plughw:Snowball"    # ALSA plughw — format conversion included
audio_device: 3                    # device index from --list-devices

Value	System	Notes
"pulse"	PulseAudio (Linux)	Recommended; configure mic via pavucontrol
"pipewire"	PipeWire (Linux)	Recommended on modern Linux
"plughw:Snowball"	ALSA (Linux)	Direct USB mic access with format conversion
"hw:2,0"	ALSA (Linux)	Raw direct access, card 2 device 0
3	Any	Device index from --list-devices
"BlackHole 2ch"	macOS	Virtual routing device

Audio preprocessing (noise reduction)

Audio preprocessing enhances a recording before transcription by filtering unwanted signal content. Noise reduction is a key preprocessing technique that removes background sound — café chatter, fan hum, keyboard clicks — while preserving speech intelligibility. Preprocessing is useful in noisy environments or when your ASR backend struggles with poor signal quality, producing errors or hallucinations. kaiku provides three noise reduction libraries, each with different strengths depending on noise type and available compute resources.

Available preprocessors

Name	Technology	Dependencies	Strengths	Weaknesses
none	—	none (default)	No overhead; baseline for clean recordings	No noise reduction; ASR errors in noisy conditions
noisereduce	Spectral subtraction	scipy only	Low CPU; scipy-only install; excellent for stationary noise (hum, AC, fans); live recording friendly	Ineffective on crowd/babble noise; may remove speech texture; limited to repeating noise patterns
pyrnnoise	Mozilla RNNoise GRU	scipy only	Handles non-stationary noise (crowd, footsteps, babble) better than spectral; learned noise patterns; no special hardware needed	16→48→16 kHz resampling overhead on 16 kHz audio; slower than noisereduce
deepfilter	DeepFilterNet3 neural	torch + Rust wheel	Best overall quality; handles mixed noise types; preserves speech naturalness and dynamics; recommended for important recordings	Highest CPU usage; heaviest dependencies; slowest option; overkill for live dictation

Choosing noise reduction by noise type and ASR backend

ASR backend noise robustness: Different backends have built-in resilience to noise due to their training data. Whisper models (including whisper.cpp and cloud APIs using whisper-large-v3-turbo like Groq) are trained on diverse YouTube videos containing natural background noise, making them inherently robust to non-stationary sounds. SenseVoice (via sherpa-onnx) is also fairly noise-tolerant. However, all backends still benefit from preprocessing in truly noisy settings (café, open office, street noise).

Preprocessing strategy by scenario:

1. Clean or quiet environment (home office, studio): Skip preprocessing (none) and let your backend's training handle any minor noise. Saves CPU and latency.

2. Steady, repeating background noise (office AC, ceiling fan, electrical hum): Use noisereduce. Spectral subtraction is highly effective on stationary patterns, adds minimal latency, and works well with any ASR backend. Ideal for live toggle recording.

3. Variable, crowd noise (meetings, café, office chatter): Use pyrnnoise. Its neural GRU approach learns diverse noise patterns better than spectral methods. Note the 16→48→16 kHz resampling overhead on standard 16 kHz recordings — acceptable for file transcription, less ideal for live recording. Pairs well with Whisper-based backends, which are already trained on YouTube's ambient noise.

4. Highest quality regardless of noise type (important interviews, archival, transcription service): Use deepfilter. The deepest neural processing delivers the cleanest output, but requires substantial CPU and time. Reserve this for offline file processing where latency and compute cost can be amortized.

Platform-specific guidance:

• Live toggle recording on CPU-constrained devices (--toggle): Prefer noisereduce over pyrnnoise or deepfilter to minimize latency and system load.
• Local whisper.cpp backend: The backend is fast but already somewhat noise-robust; consider none in quiet settings or noisereduce for steady background. Avoid deepfilter on CPU systems.
• Cloud API (Groq, OpenAI): Optional preprocessing for mildly noisy audio; use preprocessing aggressively for café-grade noise to maximize transcription quality.
• Speaker diarization (-b whisperx): Preprocessing before diarization is recommended in noisy settings — speaker separation depends on clear voice boundaries, which noise obscures.

Loudness normalisation

To complete the audio enhancement after noise reduction with any of the three preprocessors, kaiku applies a loudnorm pass (RMS → −20 dBFS, peak ceiling −0.1 dBFS) to ensure the ASR backend receives a consistently strong, unclipped signal.

Installing preprocessors

pip3 install kaiku[noisereduce]   # spectral subtraction
pip3 install kaiku[pyrnnoise]     # RNNoise GRU
pip3 install kaiku[deepfilter]    # DeepFilterNet3
pip3 install kaiku[enhance]       # all three

Preprocessor configuration

The preprocessor choice is determined using the first field of each preset spec. Override it for a single run with -p NAME.

Preprocessor usage

kaiku -p deepfilter              # denoise live recording with DeepFilterNet
kaiku -p noisereduce             # spectral denoising
kaiku -p deepfilter -i talk.mp4  # denoise video file before transcription
kaiku --test                     # also checks that configured preprocessors are available

Transcription

Flag	Description
-b NAME / --backend NAME	ASR backend to use (key under asr_backends: in config). Overrides the preset's asr_backend.
-i FILE / --input FILE	Transcribe an existing audio or video file instead of recording.
-o FILE / --output FILE	Append transcripts to file.
-l LANG / --language LANG	Language hint (ISO-639-1, e.g. fi, en). Overrides config. Omit to auto-detect.
-r / --robust	Robust mode for -i file input: split at silence boundaries, quality-check chunks, retry.
-C SEC / --chunk-duration SEC	Max chunk duration in seconds for -r/--robust mode (default: 180).
-g / --toggle	Toggle recording: first invocation starts, second stops and transcribes.
-z / --no-clipboard	Do not copy transcript (or file path) to the system clipboard; stdout and -o unchanged.

kaiku                  # record until Ctrl+C, transcribe, copy to clipboard
kaiku -l fi         # Finnish, using local whisper.cpp backend (privacy preset)
kaiku -i audio.mp3  # transcribe an existing audio file
kaiku -i meeting.mp4  # transcribe from a video file (audio extracted automatically)
kaiku -x speed -o transcript.txt   # preset + append transcript to a file

Supported input formats

• Audio: wav, mp3, m4a, ogg, flac, aac, opus, wma
• Video: mp4, mov, mkv, webm, avi, flv, mvi

Requires ffmpeg on PATH for non-WAV input. Video streams are discarded automatically; basic spectral cleaning (highpass 200 Hz + lowpass 3 kHz + loudnorm) is applied during conversion.

Language support

Use -l LANG with an ISO-639-1 code (e.g. fi, fr, de, ja) to force a specific language. Omit to auto-detect.

Whisper models distinguish between high-resource languages (English, Spanish, French, German, Portuguese, Italian, Japanese, Korean, Chinese) — which have abundant training data and near-human accuracy — and lower-resource languages with more variable results. Finnish (fi), for example, is well-handled by Whisper large-v3 and scores around 15% WER in benchmarks, which is usable for most purposes. Rare languages may require a language hint to avoid misdetection.

Backend-specific notes:

• whisper.cpp with ggml-large-v3-turbo offers the best multilingual accuracy of any local setup and is the recommended choice for non-English.
• Groq (whisper-large-v3-turbo) provides equivalent accuracy over API at very low latency — good for live recording in any supported language.
• SenseVoice (via sherpa-onnx) is exceptional for Chinese, Japanese, Korean and emotion/event detection, but its language coverage is narrower than Whisper.
• OpenAI API (whisper-1) supports the same 99 languages as the Whisper model family. Most cloud backends default to English if no hint is provided.

File and clipboard output

-o FILE appends each transcript to the specified file, with a timestamp header prepended to each entry. The file is created if it does not exist.

Continuous and chunked modes integrate with -o differently:

• In robust/chunked mode (-r), raw chunk text is appended to -o FILE as each chunk finishes (tail -f shows progress). After all chunks, the file is read back, optionally post-processed, formatted, and replaced with that final text only (no timestamped append).
• In VAD mode (--vad) and interval mode (--interval), each transcribed utterance is written immediately after the silence boundary triggers.

Clipboard size limit: when a transcript exceeds ~4 000 characters and -o FILE is specified, the file path is copied to clipboard instead of the full text.

ASR backends

ASR (Automatic Speech Recognition) converts spoken audio into text. kaiku supports several ASR backends, from cloud APIs to fully offline local inference. Speaker diarization backends (type: whisperx) live here too — they replace the regular transcription step and produce speaker-attributed output.

ASR backend configuration

ASR backends are defined under asr_backends: and referenced by name from presets or overridden per-run with -b NAME. kaiku --generate-config writes a fully annotated config with every supported backend type.

asr_backends:
  openai:
    type: api
    api_base_url: "https://api.openai.com/v1/"
    api_key: "YOUR_API_KEY"
    model_name: "whisper-1"
  groq:
    type: api
    api_base_url: "https://api.groq.com/openai/v1/"
    api_key: "YOUR_GROQ_KEY"
    model_name: "whisper-large-v3-turbo"

kaiku -b groq -i audio.wav          # use groq backend for this run
kaiku --test -b openai              # test a specific backend

Supported backend types

type	Description	Requires
api	Any OpenAI-compatible HTTP endpoint (OpenAI, Groq, SiliconFlow, xinference, etc.)	API key or local server
whisper_cpp	whisper.cpp binary via subprocess	whisper.cpp build + .bin model file
whisperx	WhisperX speaker diarization — ASR + word alignment + speaker attribution in one pass; output: [HH:MM:SS] SPEAKER_NN: text	pip3 install kaiku[diarize], HF token
mock	Fixed-response mock for testing and demos	None — no credentials needed
mock-fwd	Duration-proportional transcript mock (forward word order)	None
mock-bwd	Duration-proportional transcript mock (reverse word order)	None

Mock backends for testing

The mock backends return transcripts without making API calls or running external processes. Useful for development, demos, and CI pipelines:

asr_backends:
  demo:
    type: mock
    response: "The quick brown fox jumps over the lazy dog"
    latency_ms: 100              # Optional: simulate network delay
  mock-fwd:
    type: mock-fwd
    transcript_path: "test_data/group-2p-2.txt"   # source for word pool

kaiku -b demo -i dummy_audio.wav         # Returns mock transcript instantly
kaiku --test -b demo                     # No credentials needed
kaiku -i audio.wav -b mock-fwd           # Duration-proportional words from transcript

Local ASR: whisper.cpp vs. sherpa-onnx

Both provide fully offline, no-API-key ASR. Here is how to choose:

	whisper.cpp	sherpa-onnx (via --serve)
What it is	C++ reimplementation of OpenAI Whisper	ONNX-runtime inference with Python bindings
ASR models	Whisper family (GGML quantised)	Whisper, SenseVoice, paraformer, zipformer, and more
Language coverage	99 languages — full OpenAI Whisper training set	Varies by model; SenseVoice default covers ~50 languages (strongest for CJK); Whisper models via sherpa-onnx add 99
Multilingual quality	Best local option for European and other non-English languages; large-v3-turbo recommended	SenseVoice leads for Chinese, Japanese, Korean and handles emotion/event detection; weaker for many European languages
Python ML deps	None — single binary + model file	Yes — ONNX runtime and sherpa-onnx Python packages
Integration	Subprocess call to external C++ binary	Python-native; exposes a local HTTP API
Setup	Build C++ from source; download .bin model manually	pip3 install kaiku[vad] + kaiku --download-model
Model auto-download	No	Yes
VAD support	No	Yes (built-in via sherpa-onnx)
Dev activity	Mature, stable	Very active (k2-fsa / Next-gen Kaldi team)

When to choose whisper.cpp: Your primary language is non-English — especially European languages (Finnish, German, French, etc.) where ggml-large-v3-turbo delivers the best local accuracy of any backend. Also the right choice when you want no Python ML package dependencies: the binary and model file are self-contained, with nothing added to your Python environment.

When to choose sherpa-onnx: You are transcribing Chinese, Japanese, or Korean (SenseVoice is the stronger choice there), you want model auto-download and zero C++ build steps, you already installed kaiku[vad] (sherpa-onnx is already present), or you need VAD support or access to models beyond the Whisper family.

See whisper.cpp for build instructions and sherpa-onnx for its model zoo.

Robust long-file transcription

For long recordings, -r/--robust splits at silence boundaries, quality-checks each chunk, retries bad chunks, streams raw chunk text to -o FILE as it goes, then overwrites FILE with the post-processed, template-formatted result:

kaiku -i meeting.mp3 -r                               # chunked, quality-checked
kaiku -i m.mp3 -rC 60                                 # 60 s chunks instead of default 180
kaiku -i m.mp3 -ro transcript.txt                     # tail -f during chunks; final file is formatted output only
kaiku -i m.mp3 -l fi -o t.txt                         # fully offline, Finnish language
kaiku -i m.mp3 -rP group-restructure -T bare -o t.md  # AI meeting memo without the transcript

Long transcripts often exceed the clipboard size limit; using -o FILE is recommended.

Toggle mode

Toggle mode lets you bind a single keyboard shortcut to start and stop recording. The recording runs as a background process; the second invocation stops it, transcribes, and copies to clipboard. A desktop notification is shown on start and finish (requires notify-send on Linux).

kaiku --toggle                        # first press: start recording in background
kaiku --toggle                        # second press: stop, transcribe, copy to clipboard
kaiku --toggle                      # toggle with fully offline transcription
kaiku --toggle -P solo-restructure    # toggle → structured personal memo

Example awesome WM keybinding:

awful.key({ modkey }, "r", function()
    awful.spawn("kaiku --toggle")
end)

Requirements for toggle mode to work

Toggle mode requires a POSIX system (Linux or macOS). Windows is not supported: the recorder subprocess relies on signal.pause() (POSIX-only) and SIGTERM for graceful stop-and-write; the Windows equivalents would require reimplementing IPC with Win32 named pipes or events.

Recorder backends — Toggle mode needs audio to be captured in the background, and currently the alternative recorder backends are:

Recorder	Platform	Requirement
sounddevice (default)	cross-platform	already a dependency
arecord	Linux / ALSA only	alsa-utils system package

From these, sounddevice is used by default. If you want direct ALSA access, set recorder: arecord in config.

Local ASR server

kaiku can run a local OpenAI-compatible ASR API server backed by sherpa-onnx.

Flag	Description
--serve	Start the local sherpa-onnx ASR API server
--download-model	Download the SenseVoice model and exit
--host HOST	Server bind address (default: 127.0.0.1)
--port PORT	Server bind port (default: 8000)
--model-dir DIR	Path to ASR model directory
--num-threads N	Inference threads (default: 4)

pip3 install kaiku[vad]
kaiku --download-model                     # download SenseVoice model (~1 GB, once)
kaiku --serve                              # start server at 127.0.0.1:8000

Corresponding config backend:

asr_backends:
  sonnx:
    type: api
    api_base_url: "http://127.0.0.1:8000/v1/"
    model_name: "SenseVoiceSmall"

VAD (continuous recording)

VAD (Voice Activity Detection) classifies audio frames as speech or silence, enabling hands-free continuous transcription that triggers automatically at the end of each utterance.

Flag	Description
--vad	Continuous recording with voice activity detection. Transcribes when silence is detected after speech. Requires pip3 install kaiku[vad].
--interval SEC	Continuous recording with fixed interval (seconds).
--silence-threshold PROB	Speech probability threshold, 0.0–1.0 (default: 0.5); lower = more sensitive.
--silence-duration SEC	How long silence must last to trigger transcription (default: 1.5 s).

Continuous recording modes

Mode	Flag	Trigger	Use case
Voice Activity Detection	--vad	Silence after speech	Meetings, dictation, any unscripted speech
Fixed interval	--interval SEC	Every N seconds	Lectures, podcasts with predictable pauses

kaiku --vad -o ~/meeting.txt          # auto-transcribe when silence is detected
kaiku --interval 60 -o ~/meeting.txt  # transcribe every 60 seconds

VAD requires sherpa-onnx:

pip3 install kaiku[vad]

VAD uses the Silero VAD model (~629 KB, downloads automatically on first use). No internet connection required after the first run.

Diarization

Speaker diarization attributes each spoken segment to a speaker label, producing a transcript where every turn is tagged [HH:MM:SS] SPEAKER_NN: text. Diarization is implemented as an ASR backend (type: whisperx) — it replaces the regular transcription step entirely, so there is no double pass.

Speaker name substitution (SPEAKER_00 → real names) is intentionally left to the calling assistant or post-processor.

Flag	Description
-b whisperx	WhisperX diarization backend (type: whisperx).
-s N / --speakers N	Speaker count hint for whisperx. Ignored by other backends.

Diarization setup

pip3 install kaiku[diarize]
# Accept the pyannote licence at https://huggingface.co/pyannote/speaker-diarization-3.1
# then set your HuggingFace token:
export HF_TOKEN=hf_...

HF_TOKEN is your HuggingFace access token. You must also accept the license for pyannote/speaker-diarization-3.1 on HuggingFace before the model can be downloaded. The token is only needed on first run; the model is cached locally thereafter.

Diarization config

Define the backend among your ASR backends and reference it in a preset or pass -b whisperx:

asr_backends:
  whisperx:
    type: whisperx
    hf_token: "hf_..."       # or set HF_TOKEN env var
    min_speakers: 2          # optional hint to pyannote
    max_speakers: 6          # optional hint to pyannote

Diarization usage

kaiku -i meeting.m4a -b whisperx           # diarize: SPEAKER_NN-attributed transcript
kaiku -i meeting.m4a -b whisperx -s 3      # hint: 3 speakers (improves accuracy)
kaiku -i meeting.m4a -b whisperx -P group  # diarize + LLM meeting notes

Post-processing (with AI models)

Post-processing refines transcripts by passing them through an artificial intelligence system with custom instructions. Use this to fix transcription errors, improve grammar, condense transcripts, or restructure them into consistently formatted memos with essential information extracted. The feature is especially valuable for frequent dictators (researchers, journalists, managers) and teams with important discussions, decisions, tasks and timelines.

Flag	Description
-P NAME / --post NAME	LLM post-processor name (key in postprocessors: config) or an inline system-prompt string. Overrides the preset's post-processor.
-M MODEL / --post-model MODEL	LLM model used for post-processing. Overrides the post-processor config for this run.
-T NAME / --template NAME	Output template name from output_templates: in config. Controls what is written to clipboard / -o FILE.

Mock post-processor for testing

The mock post-processor analyzes prompts and transcripts without making API calls, returning linguistic statistics. Useful for testing post-processing workflows, demos, and CI pipelines without credentials:

postprocessor_backends:
  mock:
    type: mock
    model: Claude-Opus         # Required; becomes "Claude-Opus" in the signature

postprocessors:
  analyze:
    backend: mock
    prompt: "Enhance this transcript"

The output shows analysis of both the system prompt and the input transcript:

Prompt analyzed: longest=Enhance, shortest=this, most_frequent=enhance, lines=1, words=3, chars=24
Transcript analyzed: longest=transcription, shortest=a, most_frequent=the, lines=2, words=42, chars=245
*Yours truly, Claude-Opus
*

Each analysis includes:

• longest: longest word in the text
• shortest: shortest word in the text
• most_frequent: most common word (case-insensitive)
• lines: number of lines
• words: total word count
• chars: total character count

Usage:

kaiku -P analyze -i audio.mp3         # Analyze with mock
kaiku --test -P analyze                              # Test without credentials

Post-processors (prompt templates)

Six post-processor specs are provided in the config template as examples; each is a starting point that requires configuration:

• Setup a postprocessor_backends: entry (Ollama, Groq, Anthropic API, OpenAI, Claude Code, or any OpenAI-compatible endpoint)
• Assign that backend to the prompt (via the backend: field) or set a default with postprocessor_urgent / postprocessor_casual
• Update the context file list via context_path: to help the LLM understand context if required; Delete it if no extra context neeeded.

Examples below; see the full configuration section for all available fields.

Personal dictation prompts

Name	Purpose
solo-enhance	Improve quality, fix grammar and word choice while honoring the author's style
solo-restructure	Restructure a personal dictation into a structured memo with sections
solo-private	Like solo-restructure but defaults to a local offline model to ensure privacy

Group discussion prompts

Name	Purpose
group-enhance	Improve quality of group transcript while honoring each speaker's style
group-restructure	Restructure a group discussion into a meeting memo with summary, decisions, action items
group-private	Like group-restructure but defaults to a local offline model to ensure privacy

Tips:

• To get richer and more accurate memos from meetings and debates, use a diarization backend (-b whisperx) to attribute segments to speakers before post-processing.
• Add context files (see configuration below) to help the model better connect the dots between each individual participant and the contents discussed.

Post-processor usage examples

kaiku --toggle -P solo-enhance          # toggle → improved personal transcript
kaiku --toggle -P solo-restructure      # toggle → structured personal memo
kaiku -i meeting.m4a -b whisperx -P group-restructure  # diarize + meeting memo
kaiku --toggle -P "List action items."   # inline system prompt

Supported AI backends

Backend type	What it covers
openai_compat	Ollama (local), Groq, Anthropic API, OpenAI, any OpenAI-compatible endpoint
claude_code	Claude Code CLI — uses your CC session/subscription, no per-token billing

Post-processor backend setup

postprocessor_backends:
  ollama:
    type: openai_compat
    api_base_url: "http://localhost:11434/v1/"
    api_key: "ollama"
    model: "qwen3:14b"
  cc:
    type: claude_code
    model: "claude-haiku-4-5-20251001"

Post-processor configuration

Each prompt under postprocessors: can have the following fields:

Field	Type	Description
prompt:	string	System prompt text. Required unless extends: is used.
extends:	string	Name of another prompt to inherit from. Inherits prompt, backend, model, and context_path from parent.
extra:	string	Text appended to the inherited prompt (only with extends:).
backend:	string	Which postprocessor_backends: entry to use. Overrides inherited backend.
model:	string	Override the backend's default model for this prompt.
template:	string	Output template name from output_templates: section (default: default). Controls final output format.
context_path:	list of strings	File glob patterns to inject as context. Glob patterns are expanded and combined with inherited patterns when using extends:.

Inheritance behavior: When a prompt uses extends:, it inherits all parent fields. Child fields override parent fields, except context_path: which accumulates (both parent and child patterns are expanded and combined).

Context file formatting: Files specified in context_path: are injected into the LLM prompt with an index and clear delimiters. Each file appears with its name and visual separators, making the context easily scannable for the LLM.

Example with inheritance:

postprocessors:
  solo-base:
    backend: groq
    prompt: |
      You are a professional transcript scribe ...
    context_path:
      - "~/.kaiku/context/personal.md"     # Context file to help the LLM "read between the lines"

  solo-enhance:
    extends: solo-base            # inherits backend, prompt
    extra: |
      Also improve grammar and style ...

  solo-private:
    extends: solo-enhance        # inherits backend, prompt + extra
    backend: ollama               # override: use local model for privacy
    context_path:
      - "~/.kaiku/context/private-*.md"  # accumulates with parent's context

Note: The solo-base and group-base are not intended to be used directly. Instead, they provide definitions that are shared by other single-speaker and group discussion post-processors, respectively, through inheritance.

Output templates

Output templates control what ends up in clipboard / -o FILE. The output format depends on your template and potentially your LLM instructions — can be Markdown, ReStructuredText, plain text, or any text-based format. Two templates are shipped; select with -T NAME:

output_templates:
  raw: "{transcript}"                         # No post processing
  bare: "{result}"                            # Only the AI models output
  full: |                                     # Both the AI output and the transcript on a Markdown template
    {result}

    ---
    *Transcript from {duration_s:.0f}s recording post-processed at {datetime} with kaiku ({backend}, {prompt_name}, {model})*

    ## Original transcript

    {transcript}

kaiku -i m.mp3 -r -P group -T full      # meeting notes + full transcript appended

Available placeholders: {result} {transcript} {date} {datetime} {prompt_name} {model} {backend} {duration_s}

Set postprocessor_urgent / postprocessor_casual to apply a prompt automatically for every recording without passing -P.

Presets

Presets are atomic pipeline definitions — each specifies exactly which preprocessor, ASR backend, and post-processor to use. Pick one preset per run; kaiku handles the rest.

Format: preset_name: [preprocessor, asr_backend, postprocessor, description]

• Preprocessor — audio denoising: none, noisereduce, pyrnnoise, or deepfilter
• ASR backend — any key from asr_backends: (including diarization backends like type: whisperx)
• Postprocessor — none or any key from postprocessors:
• Description — user-facing label shown by --list / --test

All four fields are required; this keeps every processing stage explicit.

asr_backends:
  groq:
    type: api
    api_base_url: "https://api.groq.com/openai/v1/"
    api_key: "YOUR_GROQ_KEY"
    model_name: "whisper-large-v3-turbo"
  wcpp:
    type: whisper_cpp
    binary_path: "/usr/local/bin/whisper-cli"
    model_path: "/models/ggml-large-v3-turbo.bin"

presets:
  speed:     [     none,       groq,            none,  "Fast transcription with minimal processing"]
  quality:   [ deepfilter,     groq,            none,  "High-accuracy with neural denoising"]
  privacy:   [     none,       wcpp,            none,  "Fully offline, local transcription"]
  balanced:  [ noisereduce,    groq,            none,  "Good balance of speed, quality, and cost"]
  memo:      [ noisereduce,    groq,    solo-enhance,  "Transcription + LLM memo enhancement"]

kaiku --preset speed                  # record & transcribe, copied to clipboard
kaiku --preset privacy --toggle       # toggle mode with fully offline transcription
kaiku --preset quality -i audio.mp3   # transcribe file
kaiku --preset speed -b wcpp          # override backend for this run
kaiku --preset memo -P group-enhance  # override post-processor for this run

Default preset: set default_preset: speed in config to omit --preset on every invocation. Flags -b, -p, and -P still override individual stages.

default_preset: speed

Troubleshooting

Problem	Solution
Audio not captured	Run kaiku --list-devices and select a working device
Clipboard not working	Install xclip (X11) or wl-clipboard (Wayland)
API errors	Check your API key and endpoint in config
whisper.cpp errors	Run kaiku --test -b wcpp; check binary and model paths
Silent audio	Try a different audio device with --device
Video/audio format rejected	Ensure ffmpeg is installed (apt install ffmpeg / brew install ffmpeg)
Preprocessor not found	Run kaiku --test to see which are available and their install commands
Preprocessing too slow	Switch preprocessor_urgent to noisereduce or none in config
Post-processor not found	Check postprocessors: in config; name must match exactly
Post-processor backend error	Check postprocessor_backends: in config; verify API key and URL
Diarization fails	Ensure pip3 install kaiku[diarize], HF_TOKEN is set, pyannote licence accepted, and backend type: whisperx is in asr_backends:

Run kaiku --test (or kaiku --test -b <name>) to diagnose issues.

Contributing

Fork the repository and submit a pull request. Any improvements or new features are welcome! :)

Testing

Development relies on a small but powerful black-box E2E suite: the real CLI runs as a subprocess and assertions cover exit codes, stdout, stderr, and files — see tests/README.md for the full strategy, log-shape notes, and scenario index.

pytest tests/ -v

Disclaimer: Not all features have been properly tested and only on a legacy Ubuntu 20.04 environment. More testing and hardening will be done by June 2026.

License

GNU Affero General Public License v3.0. See the LICENSE file for details.

Related projects

kaiku operates within a four-stage pipeline:

[Audio capture] → [ASR / transcription] → [Post-processing] → [Output: clipboard / file]
     stage 1            stage 2                stage 3                 stage 4

The tables below cover the ecosystem at each pipeline stage and compare competing end-user tools. kaiku covers the whole pipeline in a single powerful CLI.

Audio preprocessing (noise reduction)

Audio preprocessing cleans the signal before transcription. kaiku integrates all three libraries below as optional extras (g install kaiku[enhance]); they run in a pipeline with loudness normalisation applied after cleaning.

Project	Technology	License	Best for	In kaiku
noisereduce	Spectral subtraction	MIT	Stationary noise: fans, AC, electrical hum	Yes — pip3 install kaiku[noisereduce]
pyrnnoise	Mozilla RNNoise GRU	GPL-3	Non-stationary noise: crowd, babble, footsteps	Yes — pip3 install kaiku[pyrnnoise]
DeepFilterNet	DeepFilterNet3 neural net	MIT	Best quality overall; speech naturalness; medium CPU	Yes — pip3 install kaiku[deepfilter]
RNNoise	Xiph GRU	BSD	Original Mozilla RNNoise (C library)	No — pyrnnoise wraps this at the Python layer
SpeechBrain enhance	Encoder-decoder neural	Apache-2	Research-grade speech separation and denoising	No — heavy ML framework dependency; not practical as a live preprocessor

Voice Activity Detection

VAD classifies audio frames as speech or silence, enabling automatic segment boundaries without user interaction. kaiku uses Silero VAD (bundled in sherpa-onnx) for both the --vad continuous mode and the silence-split inside --robust.

Project	License	Stars	Notes	In kaiku
Silero VAD	MIT	14k+	629 KB model; enterprise-grade; ONNX + PyTorch; auto-downloads	Yes — via sherpa-onnx in --vad and --robust
WebRTC VAD	BSD	1k+	Google's classic GMM-based VAD; very fast, lower accuracy	No — less accurate than Silero; not integrated
pyannote VAD	MIT	6k+	Neural VAD embedded in the pyannote diarization pipeline	Indirectly — activated by the whisperx ASR backend

ASR engines

ASR engines convert audio to text. kaiku is a frontend: it delegates transcription to an ASR backend, supporting two locally-run backends (whisper.cpp, sherpa-onnx) and any OpenAI-compatible HTTP endpoint for cloud or self-hosted services. The engines listed as integrated below are ones kaiku directly calls or supports as backends; the others are libraries or specialized tools that require custom wrappers to use.

Project	License	Stars	Best for	In kaiku
OpenAI Whisper	MIT	80k+	Gold standard; 99 languages; most widely reproduced	Via API (whisper-1) or indirectly through sherpa-onnx and whisper.cpp
whisper.cpp	MIT	80k+	Fully offline; best CPU performance; GGML-quantised models	Yes — type: whisper_cpp backend; subprocess call
faster-whisper	MIT	15k+	4× faster than Whisper; identical accuracy; INT8/FP16 via CTranslate2	Not directly; used internally by WhisperX and Meetily
sherpa-onnx	Apache-2	4k+	ONNX inference; multi-model-family; model auto-download; Python-native	Yes — --serve local server; type: api backend
WhisperX	BSD	13k+	Whisper + word-level timestamps + speaker diarization in one pipeline	Yes — type: whisperx backend (pip3 install kaiku[diarize])
SenseVoice	Apache-2	6k+	Emotion + language event detection; excellent CJK	Via sherpa-onnx default model; also SiliconFlow API
Vosk	Apache-2	8k+	Lightweight; 20+ languages; embedded and low-RAM devices	No — lower accuracy than Whisper family
NVIDIA Parakeet TDT	Apache-2	(NeMo)	3 380× faster than real-time; English only; GPU	No — English-only; GPU-dependent; no multilingual support
SpeechBrain	Apache-2	9k+	Research platform; fine-tuning; custom model training	No — research library, not a drop-in backend
Coqui STT	MPL-2	5k+	DeepSpeech successor; trainable on custom data	No — lower quality than Whisper; limited community activity
Kaldi	Apache-2	14k+	Enterprise/research; highly configurable; steep setup	No — complex; not a practical CLI backend

Speaker diarization

Speaker diarization labels each segment with a speaker identity ("who said what"). kaiku treats diarization as an ASR backend (type: whisperx); select it with -b whisperx or via a preset. WhisperX handles transcription and speaker attribution in one pass; pyannote.audio is used internally for speaker embedding and clustering.

Project	License	Stars	Notes	In kaiku
pyannote.audio	MIT	6k+	De-facto OSS standard; speaker embedding + clustering; requires HF token for model download	Via WhisperX (type: whisperx)
WhisperX	BSD	13k+	faster-whisper + word alignment + pyannote; all-in-one	Yes — type: whisperx backend (pip3 install kaiku[diarize])
whisper-diarization	MIT	2k+	faster-whisper + pyannote script pipeline	No — WhisperX provides equivalent functionality with an active upstream
NVIDIA NeMo	Apache-2	13k+	Fastest GPU diarization; English and enterprise focus	No — GPU-heavy; no practical CLI integration path

Desktop audio capture and transcription tools

These are end-user tools that combine audio capture, ASR, and transcript output — the closest category to kaiku itself.

Project	Type	Platform	License	Live	File	Toggle	VAD	Offline	Diarize	LLM post	Notes
kaiku (this)	CLI	Linux, macOS	AGPL-3	✓	✓	✓	✓	✓	✓	✓	Full pipeline; scriptable; multi-backend; video input
Turbo Whisper	GUI	Linux	MIT	✓	✗	✓	✗	✓	✗	✗	faster-whisper-large-v3-turbo; global hotkey; no casual mode; PPA install
Whispering	GUI/tray	Any	MIT	✓	✗	✓	✗	✓	✗	✗	Cross-platform (snap/exe); local or cloud API; minimal UI
Superwhisper	GUI	macOS, Windows, iOS	Proprietary	✓	✓	✓	✗	✓	✗	Partial	Premium dictation app; polished UX; no Linux
Meetily	Desktop app	macOS, Windows	MIT	✓	✓	✗	✗	✓	✓	✓	11.9k★; Rust backend; Ollama summaries; no Linux
Screenpipe	Agent layer	Any	MIT	✓ (always-on)	✓	✗	✓	✓	✗	Via MCP	18.6k★; ambient 24/7 recording; MCP server; not a CLI tool

Agentic voice-to-text frameworks

For AI agents a streaming architecture is often preferred: continuous listening, local wake-word detection, and real-time VAD. This allows hands-free triggering of agentic actions and conversation that feels natural. In contrast, kaiku serves as a high-precision bridge: It is currently the most complete CLI-native pipeline for when an agent (or developer) needs to process a specific audio file or a manual "push-to-talk" segment with maximum control over all the processing stages. The table below lists some of the most notable choices for AI assistants:

Project	Primary Tech	Wake Word	VAD	Agent Integration	vs. kaiku
LiveKit Agents	Python / Rust	Optional	Silero	WebSocket / WebRTC	Full Agent Framework: High-performance streaming for voice-to-voice; kaiku is a CLI tool for text generation.
Wyoming Satellite	Python	open-wakeword	Silero	Wyoming Protocol	Smart Home Focus: Designed as a background daemon for Home Assistant; kaiku is a foreground productivity tool.
Rhasspy 3	Modular (C++/Python)	Porcupine / Snowboy	WebRTC / Silero	MQTT / Unix Sockets	Deeply Modular: Can swap every component; kaiku is more integrated and opinionated for CLI users.
LocalAI	Go / C++	Yes (via API)	Yes	OpenAI-compatible API	The Server Hub: Acts as an all-in-one local API server; kaiku acts as a client that can call such servers.
Whisper Mic	Python	No	Silero	Stdout / Text stream	Simple Loop: A continuous transcription script; lacks kaiku’s noise reduction, diarization, and post-processing.
Leon	Node.js / Python	Yes	Yes	Custom SDK / Web	Full Assistant: Includes skills, memory, and UI; kaiku is a specialized "sensor" for such an assistant.

SaaS meeting assistants

Commercial cloud services that join calls automatically or process uploaded recordings. Included for context — these require trusting a third party with your audio.

Service	Platform	Bot-less	Offline	Privacy	Notable	vs kaiku
Fathom	Web (Zoom/Meet/Teams)	No	No	Cloud (US)	Free tier; calendar-integrated; good UX	Cloud-only; no file processing; no Linux CLI
Jamie	macOS, Windows	Yes	No	GDPR (EU)	Best Finnish quality; bot-free desktop app	€24+/mo; macOS/Windows only
Fireflies.ai	Web	No	No	Cloud	100 languages; CRM sync; "Ask Fred" AI queries	Cloud; BIPA lawsuit 2025; data outside EU
Granola	macOS	Yes	Partial	Cloud for AI	Calendar-integrated; note editor; bot-free	macOS only; AI processing requires cloud
Otter.ai	Web, mobile	No	No	Cloud (US)	Real-time captions; widely known	Class-action lawsuit 2025; weak Finnish
Soniox	API	—	No	Cloud (US)	Best Finnish WER (10.6%); 56 languages; developer API	API service, not an end-user tool
Krisp	App + SDK	Yes	Partial	Cloud for AI	Industry-leading noise suppression + transcription	Proprietary; subscription; not scriptable

kaiku as an open source contribution

The speech-to-text tool landscape in 2026 has a sharp divide: powerful Python libraries (faster-whisper, WhisperX, pyannote) that require programming to use, and polished end-user apps (Superwhisper, Meetily, Granola) that are macOS/Windows-only or cloud-dependent. Linux users who want local, private, keyboard-shortcut-driven transcription with the full power of the Whisper ecosystem face a gap. Turbo Whisper and Whispering address the simplest dictation case but lack file transcription, noise reduction, robustness for long recordings, and any programmable post-processing. kaiku fills this gap as a single composable CLI that exposes the full four-stage pipeline without requiring the user to write any code.

Beyond Linux, kaiku's value is its scriptability and composability. Every feature — backend, preprocessor, language, diarization, post-processor, output template — is a flag or config key. This makes it naturally callable from shell scripts, Makefiles, cron jobs, and AI coding agents: one invocation covers the full audio → transcript → structured-memo pipeline that would otherwise require stitching together three or four Python libraries. The support for both local (whisper.cpp, sherpa-onnx) and cloud (OpenAI, Groq, SiliconFlow) backends with a unified interface means the same command works offline on a laptop and in a cloud pipeline on a headless server.

The most capable competing open-source project, Meetily, is architecturally similar in ambition — local-first, offline, Whisper-backed, with LLM summaries — but is a GUI-only desktop app for macOS and Windows with no Linux support and no CLI surface. Screenpipe is a different paradigm (always-on ambient capture) rather than a competing tool. This leaves kaiku as currently the most complete open-source, Linux-native, CLI-accessible speech processing pipeline — a category with no direct competition and clear utility for developers, power users, and autonomous AI agents that need to process human speech.