Qwen3ASRSmall<T>

Models & Types · AiDotNet.SpeechRecognition.AlibabaASR

Qwen3-ASR-0.6B (Qwen Team, 2026): an LLM-based recognizer of the AuT audio encoder and the Qwen3-0.6B decoder, which names the language and then writes the transcript.

Inheritance: NeuralNetworkBase<T> → AudioNeuralNetworkBase<T> → Qwen3ASRSmall<T>

Implements: ISpeechRecognizer<T>

For Beginners

Qwen3-ASR-0.6B turns speech into a few vectors per second with its audio encoder and lets the Qwen3-0.6B language model answer a chat prompt that contains them: it writes “language English”, a marker, and then the transcript. You can force the language, and put hotwords or background text in the prompt as context.

How It Works

The report (§2) builds Qwen3-ASR-0.6B from Qwen3-0.6B, a projector and an AuT encoder (180M parameters, width 896) over 128-band Fbank features at a 12.5 Hz token rate. The release’s modeling_qwen3_asr.py fixes the encoder: 100-frame chunks of the Whisper log-mel spectrogram, three 3×3 stride-2 GELU convolutions (480 channels), a bias-free linear map to the width with sinusoidal positions restarting every chunk, pre-norm Transformer layers attending within windows of eight chunks, a layer norm, and the GELU projector to the decoder’s width. The decoder reads the projector’s output in place of the audio pads in the release’s chat template (system context; user audio between <|audio_start|> and <|audio_end|>; assistant), and answers language {name}<asr_text>{transcript}; a forced language moves its prefix into the prompt. Its interleaved multimodal rotary embedding gives audio and text the same position on all three axes, which is the ordinary rotary embedding. Training is the release’s supervised fine-tuning of every parameter on the cross-entropy of the answer and <|im_end|> after the prompt.

Example

var architecture = new NeuralNetworkArchitecture<float>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.SpeechRecognition,
    inputSize: 160000, outputSize: 151936);
var model = new Qwen3ASRSmall<float>(architecture);
var result = model.Transcribe(waveform);

Constructors

Constructor Summary
Qwen3ASRSmall(NeuralNetworkArchitecture<T> architecture, Qwen3ASRSmallOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
Qwen3ASRSmall(NeuralNetworkArchitecture<T> architecture, string modelPath, Qwen3ASRSmallOptions 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 decoder’s greedy answer, split at <asr_text> into the language it names and the transcript.
string DetectLanguage(Tensor<T> audio) The language the model names before the transcript.