vinoSpeak

Local text-to-speech on the Intel NPU, using Kokoro-82M INT8 through OpenVINO GenAI. Type a sentence or point it at a file and it reads it aloud. Nothing leaves your machine after the one-time ~160 MB model download, and OpenVINO's own telemetry is blocked.

• Speaks text, files and stdin. Markdown is cleaned up for listening, and PDFs, Word and slides work through vinoSearch's extracted text.
• Reads vinoWhisper sessions: the summary, or the transcript.
• Starts talking in about 1.5 s on the NPU, however long the text: the first sentence plays while the rest is synthesized.
• 41 Kokoro voices in US/UK English, Spanish, French, Hindi, Italian and Brazilian Portuguese, at 0.5x to 2x speed.
• Plays or saves: your speakers via PipeWire/PulseAudio/ALSA, or a WAV.
• NPU, GPU or CPU. CPU works with no driver at all.
• A window (proof of concept) to find a file with vinoSearch and listen to it, and a Python API.

Install

Python 3.11 to 3.14. With uv:

uv tool install --python 3.14 vinospeak
vinospeak "Hello."

--python 3.14 because OpenVINO 2026.4.1 has no 3.15 wheels yet. OpenVINO is pinned to the same release as vinoPop and vinoWhisper, so programs sharing one NPU never run mismatched runtimes.

From source, for the exact versions in uv.lock, hashes included. --locked fails if the lock is out of date:

git clone https://github.com/karanshukla/vinoSpeak && cd vinoSpeak
uv sync --locked
uv run vinospeak "Hello."

To put that checkout's vinospeak on your PATH with the lock's versions as constraints (uv tool install doesn't read uv.lock by itself):

uv export --locked --no-dev --no-emit-project --no-hashes -o constraints.txt
uv tool install --constraints constraints.txt .

The window isn't on PyPI, so it's always built from a checkout, with Rust. This puts it in your app menu, using the vinospeak on your PATH or in ~/.local/bin:

gui/install.sh                                             # --uninstall removes it
VINOSPEAK=.venv/bin/vinospeak gui/install.sh               # or the checkout's vinospeak

Finding files in it needs vinoSearch running. See gui/README.md.

Fedora Atomic / openSUSE Aeon: nothing to layer for vinoSpeak itself. uv brings its own Python and the wheels carry espeak-ng and libsndfile. Only the NPU driver (for --device NPU) and an audio player, if you lack one, are system packages.

No player? Install pw-play or paplay, or use --output:

Use

vinospeak "Hello from vinoSpeak."                  # play it
vinospeak "Hello." --output hello.wav              # save a WAV
vinospeak --file notes.md                          # read a file aloud
vinospeak --file notes.md --print-text             # see what it would say, no model load
some-command | vinospeak --file -                  # plain text from stdin
vinospeak --file 2026-10-27-140512.vwsession       # a vinoWhisper summary
vinospeak --file 2026-10-27-140512.vwsession --part transcript
vinospeak "Hola." --voice ef_dora --language es --speed 1.2 --device CPU

In markdown, headings, list items and table rows become sentences, links keep their text, URLs are said as their host and code blocks are skipped. Abbreviations like e.g. and Dr. don't end a sentence. vinospeak --help lists everything.

The model is fetched from Hugging Face at one pinned commit and every file is checked against its sha256 (see SECURITY.md). It lives in ~/.cache/vinospeak/, along with OpenVINO's compiled-model cache so later runs skip the NPU compile.

Check this machine

vinospeak-doctor          # --json for a bug report

Shows the OpenVINO packages and devices, the audio player it picked, and whether the cached model still matches its pins. With ovfetch on your PATH it also checks your OpenVINO version against your NPU driver. If your distro's driver has no NPU compiler (Fedora's intel-npu-driver 1.32.0), ovfetch npu install fetches Intel's into your home directory and vinoSpeak loads it from there.

Python

from vino_speak import Synthesizer

synth = Synthesizer("NPU")  # compiles once; reuse it
audio = synth.synthesize("Hello.")  # audio.samples (float32), audio.sample_rate
speech = synth.speak("A long passage.")  # plays in the background
speech.stop()  # or speech.wait()

Development

See CONTRIBUTING.md. uv run poe check runs everything CI does; none of it needs OpenVINO or an NPU.