Skip to content

TTS-VOICES.md

Latest commit

 

History

History
44 lines (34 loc) · 2.84 KB

TTS-VOICES.md

File metadata and controls

44 lines (34 loc) · 2.84 KB
Raw

Text-to-speech voice models

The "read note aloud" feature synthesizes speech using Piper voice models, pulled from the rhasspy/piper-voices collection on Hugging Face at Docker image build time (see src/StudyLife.Server/Dockerfile) - not checked into this repository. Languages without a baked-in voice fall back to the browser's own Web Speech API client-side, so the feature works for every supported UI language either way.

Baked into the default image

Language Voice License
German (de) de_DE-thorsten-low CC0 - public domain, no conditions
English (en) en_US-amy-low CC-BY-SA-4.0 - attribution + share-alike for the model itself; using it to generate speech is unrestricted

Only these two are baked into the image today - see the Dockerfile comment for why (image size trade-off). The other 19 languages with an available Piper voice (see the table below) are a possible future addition; each would need the same license check before being added.

Available but not yet baked in (Web Speech API fallback covers these today)

en_US alternate voices aside, a Piper voice exists on Hugging Face for: fr, es, it, pt, nl, da, sv, fi, el, pl, cs, sk, hu, ro, bg, sl, lv, uk, ru (21 of the app's 26 languages total, including de/en above). No Piper voice exists (as of this writing) for: hr, et, lt, mt, ga - the Web Speech API fallback is the only option for these regardless of future additions, unless a suitable model appears upstream.

Caching

TtsController caches synthesized audio in the app's existing IDistributedCache (in-memory by default, Redis in the horizontally-scaled deployment - the same instance already used elsewhere, see docs/ARCHITECTURE.md), keyed by a SHA-256 hash of the exact synthesized text + language, not the note ID. Two practical effects of that choice:

  • Editing a note produces a different hash automatically, so the cache never serves stale audio for edited content - no explicit invalidation code needed.
  • Re-reading an unchanged note (or clicking "Vorlesen" again) skips phonemization + ONNX inference entirely. Measured locally (linux/amd64, single-container): a cold request for a ~318KB/~10s note took 1.16s; the identical warm request took 10ms - about 115x faster, and byte-identical to the cold response.

Entries expire after 24h (TtsController.CacheTtl) - long enough to cover same-day re-reads, short enough that a single-container deployment's in-memory cache doesn't accumulate audio blobs indefinitely on Raspberry-Pi-class hardware.