spych 4.0.0b2

Communicate with your favorite AI model by talking to it.

Spych

Spych (pronounced "speech"): talk to your computer like it's your personal assistant — and have it talk back — without sending your voice to the cloud.

A lightweight, fully offline Python toolkit for wake word detection, audio transcription, spoken AI responses, and AI integrations. Built on faster-whisper, PvRecorder, and Kokoro.

• Can Run Fully offline: no API keys, no cloud calls, no eavesdropping (after initial setup - voice and model downloads require internet, but are cached locally for offline use thereafter)
• Multi-threaded wake word detection: overlapping listener windows so you rarely miss a trigger
• Multiple wake words: map different words to different actions in one listener
• Spoken responses: neural text-to-speech via Chatterbox Turbo (high quality, zero-shot voice cloning) or Kokoro (lightweight); agents speak their summaries aloud
• Automatic follow-up listening: when a response ends with a question, Spych listens for your reply automatically — no wake word needed
• Live transcription: continuous VAD-gated transcription to .txt and/or .srt files
• Built-in agents: for Ollama, Claude Code, Codex, Gemini CLI, and OpenCode
• Multi-agent orchestration: run several agents simultaneously under a single listener, each with its own wake words
• Personalities: named presets that bundle wake words, voice, name, and response style — e.g. --personality jarvis
• Extensible: subclass BaseResponder to build your own agents with custom wake words and logic

API Docs: https://connor-makowski.github.io/spych/spych.html

Setup

Installation

Recommended: pipx (strongly recommended)

Install Spych globally using pipx:

pipx install spych

Alternative: pip

Install using pip (requires Python 3.11+):

pip install spych

---

CLI

Once installed, spych is available as a command anywhere on your machine. You will still need to set up your respective agents before using them. See the docs below for setup instructions. Navigate to your project directory and launch any agent directly:

cd ~/my_project
spych claude

All agents and their parameters are supported as flags:

spych ollama --model llama3.2:latest
spych claude --setting-sources user project local
spych codex --listen-duration 8
spych opencode --model anthropic/claude-sonnet-4-5
spych gemini --wake-words gemini "hey gemini"

Spoken responses are enabled by default with the af_heart voice. The TTS backend (Chatterbox Turbo or Kokoro) is downloaded on first use and cached locally — all subsequent runs are fully offline.

# Uses default settings (Speaker ON, af_heart voice)
spych claude

# Change the voice
spych ollama --model llama3.2:latest --speaker_voice bm_george

# Disable spoken responses explicitly
spych claude --use-speaker false

Personality presets bundle wake words, voice, name, and response style into a single flag:

spych claude --personality jarvis
spych ollama --model llama3.2:latest --personality jarvis

A global --theme flag controls the terminal colour output and must be placed before the agent name:

spych --theme light claude
spych --theme solarized ollama --model llama3.2:latest

Available themes: dark (default), light, solarized, mono.

Live transcription is also available via the CLI:

spych live
spych live --output-path meeting --output-format srt
spych live --terminate-words "stop recording"
spych live --no-timestamps --whisper-model small.en

Multiple agents can be run by creating one terminal session per agent and setting --wake-words to be different per agent. In this way you can create 3 claude agents with different wake words.

• A Multi agent mode is also available via the CLI, but has some limitations.
• See the "Multi-agent" section below for more details.

Run spych --help or spych <agent> --help to see all available options.

---

Quick Start: Voice Agents

The fastest path from zero to voice-controlled AI. These one-liners handle everything: wake word detection, transcription, and routing your speech to the target agent.

Ollama

Talk to a local LLM entirely offline. Requires Ollama installed and running.

For this example, we'll use the free llama3.2:latest model, but any Ollama model will work. For this example run: ollama pull llama3.2:latest.

from spych.agents import ollama

# Pull the model first: ollama pull llama3.2:latest
# Then say "hey llama" to trigger
ollama(model="llama3.2:latest")

Claude Code CLI

Voice-control Claude Code directly from your terminal. Requires Claude Code installed and authenticated. See: https://code.claude.com/docs/en/quickstart. Make sure you can run claude code commands in your terminal before trying this.

Note: This can pull from your .claude folder in your user directory or from the project directory, so you can have different settings for different projects if you like.

from spych.agents import claude_code_cli

# Say "hey claude" to trigger
claude_code_cli()

Claude Code SDK

Same as above but uses the Claude Agent SDK via a subprocess worker instead of the CLI. This is great for a lightweight setup with better tool call feedback loops, but you will still need to be authenticated with the SDK and have your tools set up. See: https://platform.claude.com/docs/en/agent-sdk/overview for setup instructions.

Note: This can pull from your .claude folder in your user directory or from the project directory, so you can have different settings for different projects if you like.

from spych.agents import claude_code_sdk

# Say "hey claude" to trigger
claude_code_sdk()

Codex CLI

Voice-control OpenAI's Codex agent. Requires Codex CLI installed and authenticated. Make sure you can run codex commands in your terminal before trying this.

from spych.agents import codex_cli

# Say "hey codex" to trigger
codex_cli()

Gemini CLI

Voice-control Google's Gemini agent. Requires Gemini CLI installed and authenticated. Make sure you can run gemini commands in your terminal before trying this.

from spych.agents import gemini_cli

# Say "hey gemini" to trigger
gemini_cli()

OpenCode CLI

Voice-control the OpenCode agent. Requires OpenCode installed and authenticated. Make sure you can run opencode commands in your terminal before trying this.

from spych.agents import opencode_cli

# Say "hey opencode" to trigger
opencode_cli()

💡 Pro tip: Saying "Hey Llama" or "Hey Claude" tends to trigger more reliably than just the bare wake word.

All agents accept a terminate_words list (default: ["terminate"]). Say the word or use ctrl+c to stop the listener cleanly.

Coding Agent Parameters

Parameter	claude_code_cli	claude_code_sdk	codex_cli	gemini_cli	opencode_cli	Description
name	Claude	Claude	Codex	Gemini	OpenCode	Custom display name for the agent
wake_words	["claude", "clod", "cloud", "clawed"]	["claude", "clod", "cloud", "clawed"]	["codex"]	["gemini", "google"]	["opencode", "open code"]	Words that trigger the agent
terminate_words	["terminate"]	["terminate"]	["terminate"]	["terminate"]	["terminate"]	Words that stop the listener
model	-	-	-	-	None	Model in provider/model format
listen_duration	0	0	0	0	0	Seconds to listen after wake word (0 = VAD auto)
continue_conversation	True	True	True	True	True	Resume the most recent session
setting_sources	-	["user", "project", "local"]	-	-	-	Claude Code local settings to load
show_tool_events	True	True	True	True	True	Print live tool start/end events
use_speaker	False	False	False	False	False	Speak responses aloud via TTS
speaker_voice	"af_heart"	"af_heart"	"af_heart"	"af_heart"	"af_heart"	Voice name for spoken responses
response_style	""	""	""	""	""	Style preset or custom instruction for spoken output
spych_kwargs	-	-	-	-	-	Extra kwargs passed to Spych
spych_wake_kwargs	-	-	-	-	-	Extra kwargs passed to SpychWake

Ollama Parameters

Parameter	Default	Description
name	"Ollama"	Custom display name for the agent
wake_words	["llama", "ollama", "lama"]	Words that trigger the agent
terminate_words	["terminate"]	Words that stop the listener
model	"llama3.2:latest"	Ollama model name
listen_duration	0	Seconds to listen after wake word (0 = VAD auto)
history_length	10	Past interactions to include in context
host	"http://localhost:11434"	Ollama instance URL
use_speaker	False	Speak responses aloud via TTS
speaker_voice	"af_heart"	Voice name for spoken responses
response_style	""	Style preset or custom instruction for spoken output
spych_kwargs	None	Extra kwargs passed to Spych
spych_wake_kwargs	None	Extra kwargs passed to SpychWake

---

Summaries & Text-to-Speech (TTS)

Every agent response includes both a full response (printed to the terminal) and a short summary. The summary is always printed below long responses (over ~200 characters) so you can quickly scan what was said without scrolling. It is written to be clean prose with no file paths or special characters.

Any agent can also speak the summary aloud using the built-in neural TTS engine.

TTS Backends & Fallback

Spych uses a tiered fallback system for TTS to balance quality and performance:

1. Chatterbox (High Quality / Priority): Best for natural sounding voices and zero-shot cloning. Slower and requires more resources. Required for Python 3.14+.
2. Kokoro (Lightweight): Very fast and efficient. Ideal for edge devices (like a Raspberry Pi). Note: Not supported on Python 3.14+.

Python Runtimes

Spych supports Python 3.11+. However, due to backend dependencies, we have the following recommendations:

• Recommended: Python 3.12 — Best stability and performance using the Kokoro backend.
• Python 3.14+ — Does not support Kokoro. You must use the Chatterbox backend.

If your system's default Python is 3.14 or newer, you can force pipx to use a specific older version (if installed on your system):

# Force pipx to use Python 3.12
pipx install --python python3.12 "spych[kokoro]"

Installation Recommendations

... We recommend installing with Kokoro for most users (Python < 3.13) as it is significantly faster and uses fewer resources.

Choose Chatterbox if:

• You need high-quality voice cloning (zero-shot)
• You want to use custom voice samples (.wav files)
• You are running on Python 3.14+

Install and Run with your preferred TTS engine:

Note: If you are using python 3.13+, you will automatically install chatterbox on pip install spych

Note: If you are using python 3.12-, you will automatically install kokoro on pip install spych

By default, you will use chatterbox first if it is installed, otherwise, you will use kokoro if it is installed.

# Recommended for most users (Fast, lightweight)
pipx install "spych[kokoro]"

# For high-quality voice cloning or Python 3.14+
pipx install "spych[chatterbox]"

Enable TTS with --use-speaker (CLI) or use_speaker=True (Python). You can explicitly choose a backend with --speaker-backend:

spych claude --use-speaker --speaker-backend kokoro
spych ollama --use-speaker --speaker-backend chatterbox

When TTS is active, short responses are spoken verbatim; longer ones use the summary. If the spoken response requires user feedback, Spych automatically listens for a follow-up answer — no wake word required.

One-Shot Voice Cloning (Personalization)

One-shot cloning allows you to create a digital twin of any voice from a short audio sample. This feature is powered by Chatterbox Turbo and is not supported by the lightweight Kokoro backend.

1. Record your profile

Run the following command to record a 10-second sample of your voice. Spych will prompt you with a specific passage to read:

spych profile_my_voice --name my_voice

2. Use your custom voice

Once recorded, your voice profile is saved to the local cache and can be used by any agent. You must specify the Chatterbox backend to use custom voices:

spych claude --use-speaker --speaker-voice my_voice --speaker-backend chatterbox

Using an alternative custom voice

You can also use any .wav file as a voice profile with Chatterbox.

Simply specify the path to your .wav file instead of a profile name:

spych claude --use-speaker --speaker-voice /path/to/my_voice.wav --speaker-backend chatterbox

Note: Custom .wav profiles are only compatible with the Chatterbox backend. If you attempt to use a custom voice with Kokoro, it will fall back to using chatterbox if installed. If chatterbox is not installed, it will fall back to the default voice for kokoro.

Available Voices

The same voice names (e.g. af_heart, bm_george) work for both backends. Chatterbox uses .wav reference files for zero-shot cloning; Kokoro uses .pt voice tensors. Voice files are downloaded automatically on first use.

• Chatterbox wave voices: https://github.com/connor-makowski/spych/tree/main/voices/wave
• Kokoro pt voices (56 total): https://github.com/connor-makowski/spych/tree/main/voices/pt

American English (am_ / af_):

Voice	Gender	Grade
af_heart	F	A (default)
af_bella	F	A-
af_nicole	F	B-
am_michael	M	C+
am_fenrir	M	C+
am_puck	M	C+

British English (bm_ / bf_):

Voice	Gender	Grade
bf_emma	F	B-
bf_isabella	F	C
bm_george	M	C

---

Personalities

Personalities are named presets that bundle a wake word list, voice, display name, and response style into a single flag. They are applied as defaults — any explicit flag overrides the preset.

spych claude --personality jarvis
# equivalent to:
spych claude --name "J.A.R.V.I.S." --wake-words jarvis jarves \\
             --speaker-voice bm_george --use-speaker \\
             --response-style jarvis

Available Personalities

Name	Wake words	Voice	Style
assistant	assistant, helper, computer	af_heart	assistant — helpful, precise, informative
friend	friend, buddy, pal	af_amy	friendly — warm and simple
jarvis	jarvis, jarves, jargus, jervis	bm_george	jarvis — precise, dry wit, "sir"
pirate	blackbeard, pirate, ahoy	am_michael	pirate — pirate speak, colorful
news_anchor	bella, news anchor, anchor	af_bella	news_anchor — professional broadcast tone
robot	rob, robot	am_adam	robot — monotone, literal
caveman	er, ur, caveman, cave man	am_onyx	caveman — very simple, direct

Response Styles

The response_style parameter shapes how the LLM formats its output. Named presets:

Style	Description
assistant	Helpful and precise assistant, concise and informative
concise	Key points only, direct
friendly	Warm, approachable, simple language
military	Brevity-style, short sentences
five_year_old	Simple words, very short
fast	As brief as reasonably possible
pirate	Pirate speak, colorful
news_anchor	Professional broadcast tone
haiku	5-7-5 haiku form
shakespearean	Elizabethan English
robot	Monotone, literal
caveman	Very simple, direct
yoda	Inverted sentence structure
jarvis	J.A.R.V.I.S. from Iron Man — precise, dry wit, addresses user as "sir"

You can also pass any custom instruction string directly: response_style="Reply in exactly one sentence.".

---

Live Transcription

SpychLive continuously records from the microphone using VAD and writes the transcript to disk in real time. No wake word required — it transcribes everything until stopped.

Python

from spych.live import SpychLive

live = SpychLive(
    output_format="srt",         # "txt", "srt", or "both"
    output_path="my_transcript", # written to my_transcript.srt
    show_timestamps=True,
    stop_key="q",                # type q + Enter to stop
    terminate_words=["stop recording"],
)
live.start()

CLI

spych live                                           # writes transcript.srt
spych live --output-path meeting --output-format both
spych live --terminate-words "stop recording"
spych live --no-timestamps --whisper-model small.en

SpychLive Parameters

Parameter	Default	Description
output_format	"srt"	Output format(s): "txt", "srt", or "both"
output_path	"transcript"	Base path without extension; extensions are appended automatically
show_timestamps	True	Prepend [HH:MM:SS] timestamps to terminal and .txt output
stop_key	"q"	Key (then Enter) to stop the session
terminate_words	None	Spoken words that stop the session (detected after transcription, ~1–3s latency)
on_terminate	None	No-argument callback executed when a terminate word fires
device_index	-1	Microphone device index; -1 uses system default
whisper_model	"base.en"	faster-whisper model name
whisper_device	"cpu"	Device for inference: "cpu" or "cuda"
whisper_compute_type	"int8"	Compute precision: "int8", "float16", or "float32"
no_speech_threshold	0.4	Whisper segments with no_speech_prob above this are discarded
speech_threshold	0.5	Silero VAD probability above which a frame is considered speech onset
silence_threshold	0.35	Silero VAD probability below which a frame is considered silence during speech
silence_frames_threshold	20	Consecutive silent frames (~32ms each) required to close a segment (~640ms)
speech_pad_frames	5	Pre-roll frame count and onset confirmation threshold (~160ms)
max_speech_duration_s	30.0	Hard cap on a single segment in seconds
context_words	32	Trailing transcript words passed as initial_prompt for contextual accuracy

---

Multi-agent

Run several agents simultaneously under a single listener, each bound to its own wake words. Say "hey claude" to talk to Claude, "hey llama" to talk to Ollama — all in the same terminal session.

CLI

# Two agents, default wake words
spych multi --agents claude gemini

# Include Ollama with a specific model
spych multi --agents claude ollama --ollama-model llama3.2:latest

# Tune listen duration across all agents
spych multi --agents claude codex --listen-duration 8

Multi-agent CLI Parameters

Flag	Default	Description
--agents	(required)	One or more agent names to run: claude (claude_code_cli), claude_sdk (claude_code_sdk), codex (codex_cli), gemini (gemini_cli), opencode (opencode_cli), ollama
--terminate-words	["terminate"]	Words that stop all agents
--listen-duration	5	Seconds to listen after a wake word
--continue-conversation	true	Resume the most recent session for each coding agent
--show-tool-events	true	Print live tool start/end events
--ollama-model	llama3.2:latest	Ollama model. Only used when ollama is in --agents
--ollama-host	http://localhost:11434	Ollama instance URL. Only used when ollama is in --agents
--ollama-history-length	10	Ollama context history length. Only used when ollama is in --agents
--opencode-model	None	OpenCode model in provider/model format. Only used when opencode_cli is in --agents
--setting-sources	["user", "project", "local"]	Claude Code SDK setting sources. Only used when claude_code_sdk is in --agents

Python

Use SpychOrchestrator directly to mix any combination of responders with custom wake words.

from spych.core import Spych
from spych.orchestrator import SpychOrchestrator
from spych.agents.claude import LocalClaudeCodeCLIResponder
from spych.agents.ollama import OllamaResponder

spych_object = Spych(whisper_model="base.en")

SpychOrchestrator(
    entries=[
        {
            "responder": LocalClaudeCodeCLIResponder(spych_object=spych_object),
            "wake_words": ["claude", "clod", "cloud", "clawed"],
            "terminate_words": ["terminate"],
        },
        {
            "responder": OllamaResponder(spych_object=spych_object, model="llama3.2:latest"),
            "wake_words": ["llama", "ollama", "lama"],
        },
    ]
).start()

OrchestratorEntry Keys

Key	Required	Default	Description
responder	✓	-	A BaseResponder instance
wake_words	✓	-	Words that trigger this responder. Must be unique across all entries
terminate_words		["terminate"]	Words that stop the entire orchestrator. Merged across all entries

SpychOrchestrator Parameters

Parameter	Default	Description
entries	(required)	List of OrchestratorEntry dicts — see table above
spych_wake_kwargs	None	Extra kwargs forwarded to SpychWake (e.g. whisper_model, wake_listener_count)

---

Building Your Own Agent

Not using any of the above? No problem. Subclass BaseResponder, implement respond, and you're done. Spych handles the rest: listening, transcription, spinner UI, timing, TTS, error handling, all of it.

respond() must return an AgentResponse. Use self.format_prompt() to inject the JSON schema into your outgoing prompt and self.parse_output() to parse the result:

from spych.responders import BaseResponder, AgentResponse

class MyResponder(BaseResponder):
    def respond(self, user_input: str) -> AgentResponse:
        raw = call_my_llm(self.format_prompt(user_input))
        return self.parse_output(raw)

If you just want to echo input (e.g. for testing), construct AgentResponse directly:

from spych.responders import BaseResponder, AgentResponse

class EchoResponder(BaseResponder):
    def respond(self, user_input: str) -> AgentResponse:
        return AgentResponse(
            response=f"'{self.name}' heard: {user_input}",
            summary=f"Heard: {user_input}",
            requires_user_feedback=False,
        )

A complete working example with a custom wake word:

from spych import Spych, SpychOrchestrator
from spych.responders import BaseResponder, AgentResponse

class EchoResponder(BaseResponder):
    def respond(self, user_input: str) -> AgentResponse:
        return AgentResponse(
            response=f"'{self.name}' heard: {user_input}",
            summary=f"Heard: {user_input}",
            requires_user_feedback=False,
        )

SpychOrchestrator(
    entries=[
        {
            "responder": EchoResponder(
                spych_object=Spych(whisper_model="base.en"),
                listen_duration=5,
                name="TestResponder",
            ),
            "wake_words": ["test"],
            "terminate_words": ["terminate"],
        }
    ]
).start()

The orchestrator can also handle multiple custom agents at once, each with their own wake words. For example, you can make a translation agent that listens for "Spanish" or "German" and routes to the appropriate responder:

Note: To run this example, you will need to have Ollama running and an Ollama model that can do translations. You can use llama3.2:latest or any other model you have set up for this purpose.

from spych import Spych, SpychOrchestrator
from spych.agents import OllamaResponder
from spych.responders import AgentResponse

class Spanish(OllamaResponder):
    def respond(self, user_input: str) -> AgentResponse:
        user_input = f"Translate the following text to Spanish and return only the translated text: '{user_input}'"
        return super().respond(user_input)

class German(OllamaResponder):
    def respond(self, user_input: str) -> AgentResponse:
        user_input = f"Translate the following text to German and return only the translated text: '{user_input}'"
        return super().respond(user_input)

SpychOrchestrator(
    entries=[
        {
            "responder": Spanish(
                spych_object=Spych(whisper_model="base.en"),
                name="SpanishTranslator",
                model="llama3.2:latest",
            ),
            "wake_words": ["spanish"],
            "terminate_words": ["terminate"],
        },
        {
            "responder": German(
                spych_object=Spych(whisper_model="base.en"),
                name="GermanTranslator",
                model="llama3.2:latest",
            ),
            "wake_words": ["german"],
            "terminate_words": ["terminate"],
        }
    ]
).start()

Custom Agent Contributions

Think your agent would be useful to others? Open a PR or file a feature request via a GitHub issue. Contributions are very welcome.

---

Lower-Level API

Need more control? Use SpychWake and Spych directly.

Listen and Transcribe

Spych records from the mic and returns a transcription string.

from spych import Spych

spych = Spych(
    whisper_model="base.en",  # or tiny, small, medium, large -> all faster-whisper models work
    whisper_device="cpu",     # use "cuda" if you have an Nvidia GPU
)

print(spych.listen(duration=5))

See: https://connor-makowski.github.io/spych/spych/core.html

Wake Word Detection

SpychWake runs multiple overlapping listener threads and fires a callback when a wake word is detected.

from spych import SpychWake, Spych

spych = Spych(whisper_model="base.en", whisper_device="cpu")

def on_wake():
    print("Wake word detected! Listening...")
    print(spych.listen(duration=5))

wake = SpychWake(
    wake_word_map={"speech": on_wake},
    whisper_model="tiny.en",
    whisper_device="cpu",
)

wake.start()

See: https://connor-makowski.github.io/spych/spych/wake.html

---

API Reference

Full docs including all parameters and methods: https://connor-makowski.github.io/spych/spych.html

---

Support

Found a bug or want a new feature? Open an issue on GitHub.

---

Contributing

Contributions are welcome!

1. Fork the repo and clone it locally.
2. Make your changes.
3. Run tests and make sure they pass.
4. Commit atomically with clear messages.
5. Submit a pull request.

Virtual environment setup:

python3.11 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
./utils/test.sh