Bringing your own audio

If your app already records audio, from its own capture code, a call, a file or the network, you can use transcribe_cpp just for the transcription. Nothing on this page touches the microphone, the audio session, permissions or background handling. This is also the easy way to move an app over from a cloud speech service.

A stream of audio

transcribeStream takes audio as it arrives and transcribes it exactly like the microphone path: it finds the sentences, shows live text, and delivers results on the same text, utterances and events.

final session = await TranscribeSession.open(model);

await session.transcribeStream(
  audioChunks, // a Stream of Uint8List, Int16List or Float32List
  format: const AudioFormat(sampleRate: 48000, channels: 2),
);

It completes when the stream ends and the last sentence is delivered. Call session.stop() to end it early; audio already pushed is still transcribed.

Push audio at its natural pace to get live text. Pushed faster, for example from a file, it is transcribed as fast as the phone can, with fewer interim updates. An error on your stream ends the run, and transcribeStream throws it once what came before is delivered.

A single clip

transcribe decodes one clip in one go, with no sentence splitting and no live text:

final transcript = await session.transcribe(
  clip,
  format: const AudioFormat(sampleRate: 44100),
);
print(transcript.text);

Use it for a sentence or a short recording. A long clip runs into the model’s limits and can come back incomplete (transcript.wasTruncated); give long audio to transcribeStream, which splits it at pauses.

Audio formats

AudioFormat describes your audio, and the session converts it for you:

sampleRate	Any rate from 8 kHz to 384 kHz.
channels	Any number; they are averaged to mono. Interleaved.
sampleFormat	SampleFormat.int16 (the default) or SampleFormat.float32, little-endian.

The models take 16 kHz mono. Conversion runs off the UI isolate with a high-quality resampler, and chunks may split a sample anywhere. AudioFormat.engine is the models’ own format, which is passed through untouched.