Skip to content

sussurro.cpp (0.7.0)

Latest

Choose a tag to compare

@whispem whispem released this 28 Jun 21:17

sussurro is now multilingual. Translate between English, Spanish, French and Italian — all twelve directions — by text or by voice, fully offline.

What's new

  • Four languages, twelve directions. en/es/fr/it in every combination, no pivoting through a third language.
  • Speak or type, read or hear. The pipeline runs in any combination: text → text, text → voice, voice → text, voice → voice.
  • A voice per language. Native Piper voices for English, Spanish, French and Italian.
  • Multilingual speech-to-text. whisper transcribes in the source language you pick.
  • Desktop app. Source/target pickers, a working language swap, and native microphone capture.

Under the hood

  • The OPUS-MT / Marian models (encoder–decoder seq2seq) reimplemented from scratch on ggml.
  • Three models cover all twelve directions: one multilingual Romance↔Romance model (target chosen by a >>fra<< / >>spa<< / >>ita<< token) plus bilingual English↔Romance models.
  • Greedy + beam search, incremental KV cache, sentence splitting, f16 / q8_0 / q4_0 weights.
  • Listening via whisper.cpp; speaking via Piper voices through sherpa-onnx.
  • Runs on Apple Silicon (Metal). No Python or PyTorch at runtime.