WhisperEngine class final
Transcribes English audio with beam search.
The search matches Hugging Face's generate with num_beams beams and a
length penalty of 1, and Whisper's token suppression. fonnx used 2 beams
before this, in the ORT BeamSearch graph. Greedy decoding (1 beam) can
miss the first sentence when speech starts at 0 s.
Constructors
- WhisperEngine(WhisperOnnxBackend _backend, {int beams = 2, int maxNewTokens = 220})
Properties
- beams → int
-
final
- hashCode → int
-
The hash code for this object.
no setterinherited
- maxNewTokens → int
-
The decoder has 448 positions. The prompt uses 4.
final
- runtimeType → Type
-
A representation of the runtime type of the object.
no setterinherited
Methods
-
close(
) → Future< void> -
noSuchMethod(
Invocation invocation) → dynamic -
Invoked when a nonexistent method or property is accessed.
inherited
-
toString(
) → String -
A string representation of this object.
inherited
-
transcribe(
Float32List samples) → Future< String> -
transcribeTokens(
Float32List samples) → Future< List< int> > - Returns the text tokens of the best hypothesis.
Operators
-
operator ==(
Object other) → bool -
The equality operator.
inherited