FeengSpeak 🎙️

Give Claude Code a voice. FeengSpeak reads Claude Code's responses out loud, in real time, as they stream — fully local, no API keys, nothing leaves your machine.

Text-to-speech (TTS) voice mode for Claude Code. Hear the agent talk back while you code — hands-free, bilingual, offline.

---

Why

Claude Code talks to you in text. FeengSpeak adds the missing half: it speaks the responses aloud so you can keep your eyes on what matters, multitask, or just rest them. It hooks into Claude Code and starts reading while the response is still being generated — you don't wait for it to finish.

Built for developers: it reads the things you actually need — inline code, file paths, commands — and only skips long code blocks.

Features

• 🌎 Bilingual, automatic — detects the language of each response and reads Spanish with a Spanish voice and English with a US voice. No toggles. The language is locked per response, so the voice never flips mid-message.
• ⚡ Live streaming — reads sentence by sentence as Claude types, via the MessageDisplay hook, not after the response completes.
• 🔤 English technical terms in English — words like commit, deploy, daemon, branch are pronounced in English even inside Spanish prose (mixed-language phonemization). Editable term list.
• 👨‍💻 Dev-friendly — reads inline code, paths and commands in full; only long fenced code blocks are summarized as "code block".
• 🚀 Fast & local — Kokoro via ONNX Runtime (no PyTorch, runs on CPU). A background daemon keeps the model warm; time-to-first-audio is ~0.7s thanks to a two-stage streaming pipeline.
• 🔒 Private — 100% offline after the one-time model download. No cloud, no API keys, no telemetry.

How it works

1. Claude Code streams a response. The MessageDisplay hook feeds the text to a lightweight handler (stream_hook.py) in deltas.
2. The handler accumulates text, detects complete sentences, strips markdown (keeping inline code/paths/commands), and groups them into chunks.
3. Chunks go over a Unix socket to a daemon that synthesizes and plays them in order, without gaps — the next sentence is synthesized while the current one plays.
4. A new prompt interrupts the previous reading. A Stop hook is the fallback when live streaming is off.

Requirements

• Python 3.11+
• Linux with ALSA (aplay) for playback
• (optional) libportaudio2 for sounddevice playback + karaoke word-highlighting: sudo apt install -y libportaudio2

The TTS engine and its phonemizer (espeak-ng) are installed via pip — no other system packages required.

Install

git clone https://github.com/JuanLalo/FeengSpeak
cd FeengSpeak
./install.sh                          # venv + deps + Kokoro models (~340 MB, once)
venv/bin/python feengspeak.py setup   # installs the Claude Code hooks
feengspeak stream on                  # enable live (streaming) reading
# Restart Claude Code for the hooks to take effect.

install.sh is idempotent. Models and the venv are gitignored — a fresh clone recreates them.

Usage

feengspeak demo            # play a voice demo
feengspeak on | off        # enable / disable reading
feengspeak stream on | off # live reading while Claude types, vs reading on completion
feengspeak --voices        # list voices
feengspeak daemon-status   # daemon state
feengspeak daemon-stop     # stop the daemon (it revives on next read, reloading config)
feengspeak --voice am_onyx "any text to read"

Configuration

~/.config/feengspeak/config.json (per-user, outside the repo):

Key	Default	Description
voice	em_alex	Spanish voice (ef_dora, em_alex, em_santa).
voice_en	am_michael	US English voice (am_michael, am_adam, am_onyx, af_heart, af_nova).
auto_lang	true	Auto-detect Spanish vs English per response.
english_terms	true	Pronounce EN_TERMS words in English inside Spanish.
speed	0.93	Speech rate (lower = slower / calmer).
stream_mode	false	Live reading while typing (feengspeak stream on).
enabled	true	Global on/off (feengspeak on/off).

After changing config or code: feengspeak daemon-stop (the daemon revives and reloads).

Project layout

feengspeak.py     # CLI + daemon + synthesis + language detection
stream_hook.py    # MessageDisplay hook: live, stdlib-only, non-blocking
install.sh        # reproducible setup
requirements.txt

Credits & License

FeengSpeak is a fork of claude-voice by Null-Phnix (MIT), re-engineered for a Spanish-first, bilingual workflow: ONNX backend (no PyTorch), Spanish + English voices, mixed-language pronunciation, transcript-based extraction, and per-message language locking. See NOTICE.

Voices and engine: Kokoro / kokoro-onnx.

Licensed under the MIT License.