transcribe_cpp library
Offline, on-device speech-to-text for Flutter, over transcribe.cpp, a ggml/GGUF speech engine, with Metal acceleration on iOS.
Two entry points. ModelManager downloads and verifies a model when the app asks for one; TranscribeSession loads it and turns the microphone into text, endpointing speech hands-free.
An app with its own audio stack can skip the microphone:
TranscribeSession.transcribeStream runs the same endpointing and live
text over audio the app pushes, and TranscribeSession.transcribe decodes
a single clip. Both take any rate and channel count (AudioFormat).
Audio handed to TranscribeSession.replay must already be mono float32
at TranscribeSession.sampleRate (16 kHz), with samples in [-1, 1].
Classes
- AudioFormat
- The layout of audio handed to TranscribeSession.transcribe or TranscribeSession.transcribeStream: its rate, channel count and sample format.
- CancelToken
- Lets one caller stop waiting for a download.
- CaptureStatus
- A capture state change, with the reason when there is one.
- DownloadProgress
- How much of a model has arrived. State, not an event: a progress bar that subscribes late should draw where the download is, not wait for the next chunk.
- Dropout
- Audio the capture layer lost, and had to be skipped over.
- EndpointerConfig
-
Tuning for
Endpointer. - EndpointerEvent
-
Something the
Endpointerobserved. Times are stream-relative. - InstalledModel
- A model file that is on disk and has passed its hash check.
-
Live<
T> - A value that can be read now and watched over time.
- LiveLevel
- What the front of the pipeline is hearing, right now.
- LiveText
- The transcript so far: what is final, and what is still being spoken.
- ModelCatalog
- The models an app can offer.
- ModelDescriptor
- Everything needed to fetch, verify and identify one model file.
- ModelInfo
- Static facts about a loaded model.
- ModelManager
- Downloads, stores and tracks the models an app offers.
- NativeLog
- Controls what the engine and ggml print.
- NormalizerConfig
-
Tuning for
AudioNormalizer. - Segment
- SpeechDiscarded
- A burst of speech was thrown away for not clearing EndpointerConfig.minSpeechDuration.
- SpeechEnded
- An utterance closed and its audio is ready.
- SpeechStarted
- Speech has been confirmed — it survived EndpointerConfig.minSpeechDuration.
- Token
- TranscribeSession
- A loaded model and the pipeline around it: microphone in, text out.
- Transcript
- A complete result snapshot, fully copied out of the session.
- Utterance
-
One endpointed span of audio, ready to hand to
Transcriber.transcribe. - UtteranceResult
- A transcript for one endpointed utterance, with the timing that produced it.
- Wav
- Minimal WAV support, for capturing a session and replaying it.
- Word
Enums
- CaptureState
- What the microphone is doing.
- ModelStatus
- Where one model stands on this device.
- SampleFormat
- How each sample of pushed audio is stored. Little-endian, as on every platform the package runs on.
- SessionState
- What a TranscribeSession is doing.
- TranscribeStatus
-
Mirrors
transcribe_statusininclude/transcribe.h.
Extension Types
- ModelId
-
Names one downloadable model file, such as
parakeet-tdt_ctc-110m-Q4_K_M.
Exceptions / Errors
- DownloadCancelledException
- This caller cancelled. Bytes already received are kept for resuming.
- InsufficientStorageException
- Not enough free space to finish downloading a model.
- ModelDownloadException
- The transfer failed. Bytes already received are kept unless message says otherwise, so trying again resumes rather than restarts.
- ModelIntegrityException
- A download that cannot be installed: wrong size or wrong SHA-256.
- TranscribeException
- Thrown for any non-OK status from the native library.