models/nemoAsr/src/nemo_asr_engine library

Classes

NemoAsrEngine

Greedy decoding, as sherpa-onnx and NeMo do it.

NemoAsrOnnxBackend

Runs the ONNX graphs of a NeMo model. A TDT backend keeps the encoder output and the prediction network's state between calls.

Constants

nemoAsrLeadInSamples → const int

The quiet that withNemoAsrLeadIn ensures before speech: 250 ms.

Properties

parakeetTdtV3Blank → int

The blank of Parakeet TDT v3, its last token.

final

Functions

decodeNemoAsrText(List<String> vocabulary, List<int> tokens) → String

Joins SentencePiece tokens: ▁ starts a word. Special tokens, such as <unk> and <|nospeech|>, are skipped.

greedyCtc(Float32List logProbabilities, int vocabulary) → List<int>

Greedy CTC: the best token of each frame, without repeats and blanks. The blank is the last token.

withNemoAsrLeadIn(Float32List samples) → Float32List

Returns samples with silence before them, so that they start with at least nemoAsrLeadInSamples of quiet: 10 ms frames below -35 dBFS RMS.