MoonshineBase<T>

Models & Types · AiDotNet.SpeechRecognition.Streaming

Moonshine Base (Jeffries et al. 2024): an encoder-decoder Transformer over the raw waveform with rotary positions.

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

Implements: ISpeechRecognizer<T>

For Beginners

Moonshine Base shortens 16 kHz audio 384x with three convolutions, encodes it with Transformer layers, and a second Transformer writes the transcript token by token while reading the encoded audio. It is small and fast enough for live captions on a device.

How It Works

A three-convolution stem compresses the 16 kHz waveform 384x (widths 127, 7 and 3, strides 64, 3 and 2, with tanh and group normalization, then GELU; Figure 3), so the encoder handles audio of any length without padding. Encoder and decoder use rotary position embeddings in every layer; the encoder feed-forward networks use GELU and the decoder ones SwiGLU (Table 1). Moonshine Base is width 416 with 8 encoder and 8 decoder layers and 8 heads; the vocabulary is the Llama byte-level BPE (32,000 tokens) plus 768 special tokens, with start token 1 and end token 2. Training minimizes the decoder’s cross-entropy and transcription decodes greedily.

Example

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

Constructors

Constructor Summary
MoonshineBase(NeuralNetworkArchitecture<T> architecture, MoonshineBaseOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
MoonshineBase(NeuralNetworkArchitecture<T> architecture, string modelPath, MoonshineBaseOptions 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 decoder with greedy decoding.
string DetectLanguage(Tensor<T> audio)