Skip to content

Funktionen Offline Spracherkennung

michael edited this page Nov 3, 2026 · 4 revisions

Funktionen: Offline-Spracherkennung

Sprachmodell-Verwaltung mit Download

English Version below.

Deine Sprache wird vollständig auf deinem Gerät erkannt – ohne Internetverbindung und ohne Cloud. Das schützt deine Privatsphäre: Deine Audioaufnahmen verlassen das Gerät nie.

Wie es technisch funktioniert (kurz erklärt)

  • sherpa-onnx ist die Erkennungs-Engine, die HermiVoice eingebettet hat.
  • Zipformer-Kroko ist das verwendete Spracherkennungsmodell – kompakt und schnell direkt auf dem Gerät.

Modelle für Deutsch und Englisch

  • Deutsche und englische Modelle sind parallel installiert.
  • Je Modell fällt genau einmalig ein Download von rund 57 MB an.
  • Das aktive Modell folgt der UI-Sprache der App: Ist die Oberfläche auf Deutsch, erkennt HermiVoice Deutsch; bei Englisch entsprechend Englisch.
  • Der Wechsel zwischen den Sprachen erfolgt sofort – ohne neuen Download und ohne Neustart.

Praxis-Tipps

  • Lade die Modelle am besten im WLAN herunter.
  • Solange ein Modell für eine Sprache nicht heruntergeladen ist, kannst du in dieser Sprache keine Spracheingabe nutzen – tippe dann einfach oder wechsle die Sprache in den Einstellungen.

English

Your speech is recognized entirely on your device – no internet connection, no cloud. This protects your privacy: your audio recordings never leave the device.

How it works (in short)

  • sherpa-onnx is the speech recognition engine embedded in HermiVoice.
  • Zipformer-Kroko is the speech recognition model used – compact and fast directly on the device.

Models for German and English

  • German and English models are installed in parallel.
  • Each model requires exactly one download of about 57 MB.
  • The active model follows the UI language of the app: if the interface is German, HermiVoice recognizes German; in English mode it recognizes English.
  • Switching languages happens instantly – without a new download and without a restart.

Practical tips

  • Download the models over Wi-Fi if possible.
  • As long as a model for a language has not been downloaded, you cannot use voice input in that language – simply type instead, or switch the language in the settings.

Clone this wiki locally