[Feature] Add Automatic Speech Recognition (ASR) support

[eav-solution]

Add support for Automatic Speech Recognition (ASR) models and pipelines to enable speech-to-text transcription capabilities.
Models: https://huggingface.co/collections/litert-community/asr

[DenisovAV]

Hi @eav-solution — thanks for the suggestion. The ASR models in
litert-community use a different runtime than what flutter_gemma
currently supports (raw LiteRT interpreter + per-architecture decode

• audio capture, not LiteRT-LM Engine API) — adding to the roadmap.

Sasha

[DenisovAV]

Hi @eav-solution — this shipped!

On-device ASR is now available in the new flutter_gemma_speech package (part of the 1.4 release). It runs the litert-community ASR models you linked directly on the raw LiteRT interpreter (not the LiteRT-LM Engine API) — exactly the runtime gap I mentioned above.

• moonshine-tiny (raw 16 kHz PCM) is wired and verified end-to-end today. The pipeline is model-selectable (SttModelType).
• Whisper and Parakeet already pass LiteRT op-coverage — the runtime runs them; they just need their log-mel frontend profile wired (currently UnimplementedError), so those are a follow-on.
• Usage: initialize(sttBackends: [LiteRtSttBackend()]) → installStt()…ofType(SttModelType.moonshine).install() → getActiveStt() → transcribe(pcm16kMono).
• Native only for now (Android / iOS / macOS / Windows / Linux); Web is a follow-on.
• Docs: https://fluttergemma.dev/docs/speech · Package: https://pub.dev/packages/flutter_gemma_speech

Give it a try and let me know how it goes — closing this as shipped, but feel free to reopen or ping if you hit anything.

Sasha