waken-voice

Voice Source and Output for Waken — "nginx for AI agents." Unlike waken-claude/waken-gemini/waken-copilot, this package doesn't wrap an AI backend at all: it wraps a channel. VoiceSource turns an audio file into an Event; VoiceOutput turns a Response back into audio. Both talk to a pluggable backend — OpenAI (Whisper + TTS) by default, or swap in Groq (Whisper-backed transcription) and/or gTTS (free, no-API-key speech synthesis).

What this is NOT: a live microphone listener

Continuous microphone capture with voice-activity-detection to segment utterances is a real, hard, hardware/OS-dependent problem — and not one an autonomously-built package can honestly claim to have solved without a real microphone and a human speaking into it to verify against. So VoiceSource doesn't try. Instead it watches a directory for audio files appearing on disk, exactly like the built-in FilesystemSource watches for any file — dropped there by a phone system, a Telegram/Slack voice-message webhook, a push-to-talk recorder, or anything else. That's a small, honest, fully testable scope that composes cleanly with whatever actually captures audio, which is the same "small, composable pieces" philosophy Waken itself is built on.

Install

pip install "waken-voice[openai]"   # default: Whisper + TTS via OpenAI
pip install "waken-voice[groq]"     # transcription via Groq instead
pip install "waken-voice[gtts]"     # synthesis via gTTS instead (free, no key)

Each provider is an extra, not a hard dependency — install only the ones you use. openai, groq, and gtts can all be combined in one install (e.g. pip install "waken-voice[groq,gtts]" for a fully OpenAI-free setup).

• OpenAI: set OPENAI_API_KEY in the environment, or pass api_key=... to OpenAIWhisperTranscriber/OpenAITTSSynthesizer.
• Groq: set GROQ_API_KEY, or pass api_key=... to GroqWhisperTranscriber.
• gTTS: no API key — it calls Google Translate's public (undocumented) TTS endpoint directly.

Usage

Default (OpenAI for both directions — no transcriber/synthesizer arguments needed):

from waken import Runtime
from waken_openai import OpenAIAdapter
from waken_voice import VoiceSource, VoiceOutput

runtime = Runtime()
runtime.target("openai", OpenAIAdapter())
runtime.source("voice-in", VoiceSource(watch="./voice-inbox", target="openai"))
runtime.output("voice-in", VoiceOutput(output_dir="./voice-outbox"))
runtime.run()

Groq for transcription, gTTS for synthesis:

from waken import Runtime
from waken_openai import OpenAIAdapter
from waken_voice import VoiceSource, VoiceOutput, GroqWhisperTranscriber, GTTSSynthesizer

runtime = Runtime()
runtime.target("openai", OpenAIAdapter())
runtime.source(
    "voice-in",
    VoiceSource(
        watch="./voice-inbox",
        target="openai",
        transcriber=GroqWhisperTranscriber(),  # reads GROQ_API_KEY
    ),
)
runtime.output(
    "voice-in",
    VoiceOutput(output_dir="./voice-outbox", synthesizer=GTTSSynthesizer(lang="en")),
)
runtime.run()

Drop a .wav/.mp3/.m4a/.ogg file into ./voice-inbox; VoiceSource transcribes it and dispatches an Event(payload={"prompt": <transcript>}) to the "openai" target. runtime.output(name, ...) registers an Output under a name resolved by event.source (or an explicit event.output) — so registering VoiceOutput under "voice-in", the same name as the source above, means a Response to a voice-originated Event gets spoken back by default; see docs/api-spec.md §3 and §9 in the main waken repo for the full delivery-resolution rule.

Providers

VoiceSource takes a Transcriber (anything with an async transcribe(path) -> str method); VoiceOutput takes a Synthesizer (anything with an async synthesize(text) -> bytes method). Both are plain Protocols, not base classes — write your own by matching the method signature, same name-keyed-object philosophy as runtime.target("claude", ClaudeAdapter()) in core Waken.

**client_kwargs forwards to the provider's own async client (openai.AsyncOpenAI(...) / groq.AsyncGroq(...)); **gtts_kwargs forwards to gtts.gTTS(...) (e.g. slow=True). Each class imports its SDK lazily on construction, so you only need the extra actually installed for whichever class you use.

gTTS is worth knowing the shape of before relying on it: it's a thin wrapper around Google Translate's public but undocumented TTS endpoint, not an official API with an uptime or rate-limit contract, and it offers a fixed accent-per-language voice rather than named-voice selection like OpenAI's alloy/nova/... It's genuinely free and requires no signup, which is the whole appeal — just don't reach for it if you need guaranteed availability or voice control.

VoiceSource

VoiceSource(
    watch,                  # directory to poll for new audio files
    target,                 # name of the registered Target to dispatch to
    interval=1.0,           # poll interval, seconds
    source_name="voice",
    transcriber=None,       # a Transcriber; defaults to OpenAIWhisperTranscriber()
)

Polls watch every interval seconds for new files, same mechanism as the built-in FilesystemSource (see waken/plugins/sources/filesystem.py) — no OS-level file-watching dependency, files present before start() are the baseline and never fire. Only .wav, .mp3, .m4a, and .ogg files are transcribed; anything else is ignored silently, since handing a non-audio file to a transcriber is a guaranteed, pointless error rather than a useful attempt.

VoiceOutput

VoiceOutput(
    output_dir,          # directory to write synthesized audio into
    synthesizer=None,    # a Synthesizer; defaults to OpenAITTSSynthesizer()
    play=True,           # attempt local playback after writing the file
)

Writes {event.event_id}.mp3 under output_dir. A Response with no text delivers nothing.

Playback (play=True, the default) is best-effort and untested against real audio hardware or OS combinations. _play shells out to a platform player (afplay on macOS, paplay on Linux, an unverified fallback elsewhere) via asyncio.create_subprocess_exec, after the file is already safely written to disk — a missing or failing player is logged, not raised. This is the one piece of the package that genuinely can't be verified by an agent with no speakers to listen to; treat it as a starting point for your own environment, not a guarantee. Set play=False and hook your own playback (or upload/streaming path) onto the written file if you need something verified for your setup. (Same honest-about-limits spirit as waken-copilot's README regarding its own unverified pieces.)

Development

git clone https://github.com/WakenHQ/waken-voice
cd waken-voice
pip install -e ".[dev]"
pytest

dev pulls in openai, groq, and gtts together so the full suite can run. Tests mock each provider's SDK client (openai.AsyncOpenAI, groq.AsyncGroq, gtts.gTTS) and VoiceOutput._play entirely — no real network access, API key, or audio hardware is required to run them.

License

MIT