CIFEncoder<T>

Models & Types · AiDotNet.SpeechRecognition.ConformerFamily

CIF (Continuous Integrate-and-Fire) encoder for soft monotonic alignment in ASR.

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

Implements: ISpeechRecognizer<T>

For Beginners

CIF provides a soft, monotonic alignment mechanism between encoder frames and output tokens. Each encoder frame produces a firing weight; when cumulative weight exceeds a threshold (1.0), the accumulated weighted features “fire” as an output token…

How It Works

References:

  • Paper: “CIF: Continuous Integrate-and-Fire for End-to-End Speech Recognition” (Dong & Xu, 2020)

CIF provides a soft, monotonic alignment mechanism between encoder frames and output tokens. Each encoder frame produces a firing weight; when cumulative weight exceeds a threshold (1.0), the accumulated weighted features “fire” as an output token representation. This enables non-autoregressive decoding (like Paraformer) while maintaining monotonic alignment.

Example

// Create a CIF encoder model for soft monotonic alignment ASR
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.SpeechRecognition,
    inputHeight: 16000, inputWidth: 1, inputDepth: 1, outputSize: 5000);
var model = new CIFEncoder<double>(architecture);
            
// Or load a pre-trained ONNX model for CIF-based encoding
var onnxModel = new CIFEncoder<double>(architecture, "cifencoder.onnx");

Constructors

Constructor Summary
CIFEncoder(NeuralNetworkArchitecture<T> architecture, CIFEncoderOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates a CIF encoder in native training mode.
CIFEncoder(NeuralNetworkArchitecture<T> architecture, string modelPath, CIFEncoderOptions options) Creates a CIF encoder in native training mode.

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)
TranscriptionResult<T> Transcribe(Tensor<T> audio, string language, bool? includeTimestamps) Transcribes audio using CIF-based soft monotonic alignment.
string DetectLanguage(Tensor<T> audio)