FrontPocket
June 4, 2026 · View on GitHub
FrontPocket provides a front-end to Kyutai Labs Pocket TTS, including the ability to read text from the clipboard, from a text file and passed directly on the CLI. Features include the ability to pause, resume, move back and forward in the spoken text and change playback speed.
It is a low-latency, daemon-based text-to-speech system, developed and tested under Linux. FrontPocket loads the TTS model once at startup and streams audio sentence by sentence, so there is minimal delay between sending text and hearing it spoken.
Subsequent sentences are generated in advance and previous sentences are cached allowing instantaneous movement backwards and forwards a few sentences.
frontpocket_server.py is intended to run in the background as a systemd service. You can also run it from the CLI for interactive use — recommended while getting it dialled-in before committing to running it as a service. The server is controlled by the included lightweight CLI client.
You could set hotkeys to run the CLI client passing it parameters to play, pause, change speed, etc. e.g. Ctrl-Shift-S could trigger speaking of the clipboard text.
A Qt6-based toolbar frontpocket_toolbar.py provides a UI to drive the CLI client — play clipboard text, move forward and back, and change voices and speed. Right-click to see the speed and voice selections.
Compact Toolbar:
Right click to expand toolbar:
Note: Developed and tested under Debian Linux. macOS/Windows should work but has not been tested. Please test and submit a PR for any needed fixes.
Inspiration for this project comes from Kokorodoki, which provides a similar feature set for Kokoro TTS.
Much thanks to the very smart people at Kyutai Labs for their beautiful model and helpful reference code. Their work is where the real magic happens. https://github.com/kyutai-labs/pocket-tts
Features
- Low latency — model is pre-loaded; audio begins within seconds of sending text
- Chunk-ahead generation — the next several sentences are generated in the background while the current one plays
- Lookbehind cache — previously played sentences are cached for instant
!backnavigation - Multiple voices — switch voices on the fly; built-in voices and custom
.safetensorsembeddings supported - Speed control — pitch-preserved speed adjustment via pyrubberband
- Pause / resume — resume from exactly where you paused, even after changing voice or speed
- Skip forward / back — move through sentences instantly; previously played sentences are cached
- Interrupt — inject an urgent TTS message mid-playback, then resume automatically
- Clipboard-first — default input is the system clipboard; also accepts inline text and text files
- Auto model reload — detects bad TTS generation and reloads the model automatically
- systemd ready — runs as a proper systemd user service with automatic restart on failure
- User-friendly Polish — proper handling of numbers, currencies, hyphens, acronyms. Workarounds for words that the Pocket-TTS model has difficulty with.
- Multilingual — sentence segmentation supports English, German, French, Spanish, Italian, Russian, Polish, and more. Tested with English. Full multilingual support coming in a future version.
- UI Toolbar - Because we don't always want to be in the CLI.
How It Works
FrontPocket has three components:
| Component | File | Role |
|---|---|---|
| Server | frontpocket_server.py | Loads the model, listens on a TCP socket, plays audio |
| Client | frontpocket_client.py | Sends text or commands to the server |
| Toolbar | frontpocket_toolbar.py | Provides a UI to talk to the client |
The server and client communicate over a local TCP socket (default port 5562). The client is fire-and-forget — it sends a message and exits immediately.
Installation
See INSTALL.md for full manual setup instructions including systemd service configuration.
Starting in FrontPocket v1.4 frontpocket_installer.sh is provided. This script simplifies installation on Linux environments. Tested on Debian.
Download the script, chmod +x frontpocket_installer.sh, ./frontpocket_installer.sh
The script will download the project and perform the same install steps as the manual install. This needs more testing, especially on non-Debian systems. A log file is created in ~/FrontPocket which may help with troubleshooting and issue reporting.
Client Usage
fp [text] [options]
Input
| Command | Description |
|---|---|
fp | Speak clipboard contents (default) |
fp "Some text" | Speak inline text |
fp --file article.txt | Speak contents of a text file |
Playback Control
| Command | Short | Description |
|---|---|---|
fp --pause | Pause playback | |
fp --resume | Resume from where you paused | |
fp --next | Skip to next sentence | |
fp --back | Go back one sentence |
Why no --stop? Just use pause and don't resume.
Settings
| Command | Description |
|---|---|
fp --voice mary | Change voice (takes effect immediately) |
fp --speed 1.5 | Change speed (0.5–3.0, takes effect immediately) |
Interrupts & Status
| Command | Description |
|---|---|
fp --interruptwith "text" | Pause, speak the text, resume |
fp --interruptwith alert.txt | Same, but read text from a file |
fp --status | Speak current voice, speed, and playback state |
Other Options
| Option | Description |
|---|---|
--ping | Check server is reachable (exit 0 = up, exit 1 = down) |
--list-voices | Print all voices configured in frontpocket.ini |
--version | Print FrontPocket version and exit |
--port PORT | Connect to a non-default server port |
--host HOST | Connect to a non-default server host |
--quiet | Suppress all client output |
Configuration
FrontPocket is configured via frontpocket.ini. The server looks for it next to frontpocket_server.py, then in the current working directory. When installed as a service, a symlink is created from ~/FrontPocket/ to ~/.config/FrontPocket/frontpocket.ini. Perform your changes in ~/.config/FrontPocket/frontpocket.ini.
Key settings
[settings]
default_voice = alba
default_speed = 1.0
port = 5562
lookahead_chunks = 5
language = en
log_level = INFO
interrupt_sound = ~/FrontPocket/sounds/notification.wav
[voices]
alba = alba
mary = ~/FrontPocket/voices/mary.safetensors
See the fully commented frontpocket.ini for all available options.
Voices
FrontPocket uses pocket_tts for TTS.
Built-in voices are referenced by name in frontpocket.ini:
alba = alba
Custom voices use .safetensors embedding files:
mary = ~/FrontPocket/voices/mary.safetensors
Hugging Face voices can be referenced directly:
expresso = hf://kyutai/tts-voices/expresso/ex01-ex02_default_001_channel2_198s.wav
Commands Reference (Server Protocol)
Text sent to the server without a ! prefix is spoken. Commands are prefixed with !:
| Command | Alias | Description |
|---|---|---|
!pause | !p | Pause playback |
!resume | !r | Resume playback |
!next | !n | Skip to next chunk |
!back | !b | Go back one chunk |
!voice <name> | Change voice | |
!speed <value> | Change speed | |
!interruptwith <text> | !i | Interrupt with text, then resume |
!status | Speak current voice, speed, and state | |
!ping | No-op — used to verify server is reachable |
Logging
| Level | What you see |
|---|---|
ERROR | Fatal errors only |
INFO | Chunk generation, playback, commands received (default) |
DEBUG | Everything above plus chunk text, socket messages, and per-chunk .txt files in debug_dir |
Set log_level in frontpocket.ini or override at launch:
python3 frontpocket_server.py --log-level DEBUG
When running as a service, view logs with:
journalctl --user -u frontpocket -f
Dependencies
| Package | Purpose |
|---|---|
pocket_tts | TTS engine |
pysbd | Sentence boundary detection |
pyrubberband | Pitch-preserved speed adjustment |
sounddevice | Audio playback |
numpy | Audio array handling |
scipy | WAV file reading (interrupt sound) |
rubberband-cli | System package required by pyrubberband |
pyqt6 | UI for the toolbar |
pyperclip | Windows clipboard support (optional) |
python3-venv | venv module for python3 |
emoji | Identify and map emojis in text |
num2words | 21 -> twenty one |
pronouncing | Word list |
License
MIT