ParakeetTDT<T>

Models & Types · AiDotNet.SpeechRecognition.NeMo

Parakeet-TDT: NVIDIA NeMo’s Fast Conformer with a token-and-duration transducer (Xu et al. 2023).

Inheritance: NeuralNetworkBase<T> → AudioNeuralNetworkBase<T> → ParakeetTDT<T>

Implements: ISpeechRecognizer<T>

For Beginners

Parakeet-TDT turns 16 kHz audio into an 80-band log-mel spectrogram, shortens it 8x with depthwise-separable convolutions, and encodes it with Conformer blocks, each mixing self-attention (long-range context) with a convolution (local patterns). A small LSTM remembers the words written so far and a joint network combines it with each audio frame to pick the next token or “nothing yet”.

How It Works

The encoder is Fast Conformer at parakeet-tdt-0.6b’s size: 8x depthwise-separable subsampling of 256 channels, 24 blocks, width 1024, 8 heads and a 9-wide depthwise kernel (Rekesh et al. 2023, section 3.1). The decoder is a TDT: a 640-unit LSTM prediction network and a 640-wide joint network whose output also predicts each emission’s duration among 0 to 4 frames, trained with the TDT loss and decoded by skipping the predicted frames. SpecAugment masks the features in training (section 3.1). The decoder trains with the token-and-duration transducer loss and decodes greedily. The paper’s blocks use relative positional encoding; these use rotary position encoding, the relative scheme ConformerBlockLayer provides.

Example

var architecture = new NeuralNetworkArchitecture<float>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.SpeechRecognition,
    inputSize: 16000, outputSize: 34);
var model = new ParakeetTDT<float>(architecture);
var text = model.Transcribe(waveform).Text;

Constructors

Constructor Summary
ParakeetTDT(NeuralNetworkArchitecture<T> architecture, ParakeetTDTOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
ParakeetTDT(NeuralNetworkArchitecture<T> architecture, string modelPath, ParakeetTDTOptions options)

Properties

Property Summary
IReadOnlyList<string> SupportedLanguages { get; }
bool SupportsStreaming { get; }
bool SupportsWordTimestamps { get; }

Methods

Method Summary
IReadOnlyDictionary<string, T> DetectLanguageProbabilities(Tensor<T> audio)
IStreamingTranscriptionSession<T> StartStreamingSession(string language)
ModelMetadata<T> GetModelMetadata()
Task<TranscriptionResult<T>> TranscribeAsync(Tensor<T> audio, string language, bool? includeTimestamps, CancellationToken cancellationToken)
Tensor<T> ForwardForTraining(Tensor<T> input)
TranscriptionResult<T> Transcribe(Tensor<T> audio, string language, bool? includeTimestamps) Transcribes audio: the encoder and the transducer with greedy decoding.
string DetectLanguage(Tensor<T> audio)