Repository navigation
Add standalone Parakeet TDT v3 utterance transcription - #115
Merged
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Add a standalone Parakeet TDT v3 INT8 utterance API for callers that own segmentation. It accepts bounded 16 kHz mono audio and returns text plus decoder-emission word times without constructing a VAD/TTS pipeline. Calls are serialized, inputs are validated, and close is idempotent.
The three model files use a separate cache, immutable revision and SHA-256 verification. ModelDownloadWorker can fetch this set explicitly without changing existing default model selections. The pinned core includes corrected Nemotron features, Android scratch allocation lifetime and failed-load cleanup (soniqo/speech-core#167).
The model supports 25 languages with automatic detection and has no language-forcing prompt. Emission times are not forced alignment or speaker evidence. Calls are bounded to 30 seconds; callers must evaluate their own segmentation and latency before switching a production stream.
Validation: SDK, app and control-demo JVM suites pass. Consumer arm64 Android runs cover repeated utterances, complete recordings, Auto/selected/Auto switching, word retention, speaker flushing and model failure/retry cleanup. The selected SDK connectedAndroidTest suites pass 12/12 tests on an isolated arm64 API 35 emulator: Parakeet batch, Nemotron streaming, Sortformer and ReDimNet. Added device tests cover authored synthetic speech, repeatability, bounded word times, closed handles and incomplete bundles. Consumer pipeline checks cover VAD, speaker flushing and saved-voice recognition. No private recordings or transcripts are included. Physical microphone and performance qualification remain consumer release checks.