Zum Hauptinhalt springen

Phonon TTS Smart-Home-Integration, kostenlos und Open Source

Phonon TTS-Integration für Gladys Assistant

Offline text-to-speech: Gladys speaks with Gradium's Phonon voices, on your own hardware.

Auf GitHub ansehenAlle Integrationen
Info

Die Dokumentation dieser Integration wird von ihrem Autor geschrieben und ist bisher nur auf Englisch verfügbar.

Phonon TTS gives Gladys an offline voice: the texts to speak (scene "Play a notification" announcements) are synthesized on your Gladys machine by Gradium's Phonon models, with no cloud service or subscription. Once the model is downloaded, no internet connection is needed.

Prerequisites​

  • Gladys 5.3 or later (pluggable voices).

  • Gladys in host network mode (standard installation): the speakers (Sonos, Google Cast, AirPlay…) download the audio from an address of your local network.

  • About 1 GB of memory and 150 MB of disk free for the speech engine. A Raspberry Pi 4 or 5 is fine.

  • For the Phonon 40m models (every language): a free Hugging Face account.

    1. Open the model page, for instance gradium/phonon-40m-fr, and accept its terms of use (approval is instant).
    2. Create a read token at huggingface.co/settings/tokens.

    The Phonon 31m model (English only) needs no account.

The Phonon 40m models are licensed CC BY-NC 4.0 (non-commercial use); Phonon 31m and the catalog voices are licensed Apache 2.0.

Configuration​

Right after installation, the integration does not speak: no model is selected by default and nothing is downloaded.

  1. Open the Configuration tab of the integration.
  2. Choose the model: a 40m model in your language (the best choice), the multilingual model (then choose its language), or Phonon 31m for English without an account.
  3. For a 40m model, paste your Hugging Face token. It is only used for the download and never sent to the speech engine.
  4. Choose the voice: the model's own voice, or one of the 102 voices of the Phonon catalog. Listen to them first on the voice demo page, which shows the exact name to pick in the list. Every voice works with every model, but sounds best in its own language.
  5. Save. The status at the top of the screen follows the download (about 75 MB), then the engine startup, until Ready.
  6. Make Phonon TTS the voice of Gladys: Integrations › Artificial Intelligence › Voice, choose Phonon TTS.

Your "Play a notification" scenes now speak with Phonon.

Only the selected model and voice are kept on disk: switching model or voice deletes the previous ones. While a new model downloads, the current voice keeps speaking.

Custom voice (Gradium Studio)​

A voice designed or cloned in Gradium Studio works too:

  1. Create an API key in Gradium Studio, paste it into Gradium API key, and save.
  2. Run the List my Gradium Studio voices action: it shows the ID of each of your voices.
  3. Choose Custom voice in the voice list, paste the ID into Custom voice, and save.

You can also enter the https URL of a Phonon voice file (.safetensors) there. The voice is downloaded once.

Advanced settings​

  • Speech rate: slower, normal or faster.
  • Model weights: q8 (recommended) is smaller and faster on a CPU; f32 is larger and slower.

Actions​

  • Test the voice: synthesizes a sentence (yours, or a sample one) and reports the length of the generated speech and the computing time.
  • List my Gradium Studio voices: see above.
  • Restart the speech engine: retries the download and restarts the engine, for instance after a network outage.

Limitations​

  • The voice assistant answers in text with Phonon TTS: its page is served over HTTPS, and browsers refuse to play audio served over plain HTTP on the local network. Scene announcements speak normally.
  • Gladys waits 30 seconds at most for a synthesis. On a modest machine, a very long text is cut and only its beginning is spoken.
  • One language per model (the model's, or the one chosen for the multilingual model): a text in another language is read with the model's accent.

Troubleshooting​

  • "No model selected": choose a model and save.
  • "Gated model" / "Access denied": accept the terms on the Hugging Face page of the model with the account of the token, check the token, then save again.
  • "Cannot reach Hugging Face": Gladys needs the internet for the first download. Check the connection, then run Restart the speech engine.
  • The engine stops while starting: check the logs of the phonon container in the Supervision block (not enough memory, unsupported CPU…).
  • The speakers play nothing: check that Gladys runs in host network mode and that Phonon TTS is the voice selected in Integrations › Artificial Intelligence.
  • The logs of the integration (Supervision block) detail every step.

Report an issue at github.com/bertrandda/gladys-phonon.

Konfigurationseinstellungen​

Diese Einstellungen fragt Phonon TTS in seinem Konfigurationsbildschirm in Gladys ab.

EinstellungTypPflichtfeldBeschreibung
Phonon modelsectionNeinNothing is downloaded and Gladys cannot speak until you choose a model and save. The 40m models need a free Hugging Face account: open the model page, accept its terms (approval is instant), then create a read token. They are licensed CC BY-NC 4.0 (non-commercial use). Phonon 31m (English) is Apache 2.0 and needs no account. Once downloaded, everything runs offline.
ModelselectJaEach 40m model speaks one language. The multilingual one speaks all five: pick its language below.
Language (multilingual model only)selectNeinHow numbers, dates and symbols are read out. Ignored by the single-language models.
Hugging Face tokensecretNeinRead token of the account that accepted the model's terms. Only used to download a 40m model, never sent to the speech engine.
VoicesectionNeinChoose the model's own voice or one of the voices of the Phonon catalog: listen to them on the demo page first. Only the selected voice is downloaded. Every voice works with every model, but sounds best in its own language. You can also use a voice you designed or cloned in Gradium Studio.
VoiceselectNein
Custom voicestringNeinUsed when the voice is "Custom voice": the ID of a voice designed or cloned in Gradium Studio (the "List my Gradium Studio voices" action below shows them), or the https URL of a Phonon voice file (.safetensors).
Gradium API keysecretNeinOnly for a custom voice given by its ID: create a key in Gradium Studio. The voice is downloaded once, then works offline.
AdvancedsectionNeinThe defaults suit most installations.
Speech rateselectNein
Model weightsselectNeinq8 is smaller and faster on a CPU, with the same voice quality for most uses.

So installierst du Phonon TTS in Gladys​

  1. Öffne in Gladys Integrationen: Phonon TTS erscheint im Katalog neben den nativen Integrationen, mit einem Community-Badge.
  2. Klicke auf Installieren. Gladys lädt das Docker-Image (ghcr.io/bertrandda/gladys-phonon:1.0.0) herunter, startet es in einer vom Kern isolierten Sandbox und erzeugt die Oberfläche der Integration (Geräte, Erkennung und Konfiguration).
  3. Öffne den Bildschirm Konfiguration der Integration, fülle die Einstellungen aus und speichere.
  4. Du kannst sie auch direkt über die URL ihres Repositorys installieren: https://github.com/bertrandda/gladys-phonon.

Phonon TTS benötigt Gladys >=5.3.0. Der Katalog in Gladys wird stündlich aktualisiert, eine neue Version ist also spätestens eine Stunde nach ihrer Veröffentlichung verfügbar.

Du nutzt Gladys noch nicht? Es ist kostenlos und Open Source: folge der Installationsanleitung, um loszulegen.

Über externe Integrationen​

Phonon TTS ist eine externe Integration: eine Community-Integration, die als Docker-Container verpackt und auf GitHub veröffentlicht wird. Gladys installiert sie mit einem Klick und führt sie in einer vom Kern isolierten Sandbox aus. Sie wird von bertrandda veröffentlicht und gepflegt, nicht vom Gladys-Kernteam.

Abonniere den Newsletter von Gladys Assistant

Ein paar E-Mails im Monat zu neuen Releases und Neuigkeiten aus dem Projekt. Verschickt von Pierre-Gilles Leymarie, dem Gründer des Projekts. Jederzeit abbestellbar 🙂
Zuletzt aktualisiert am