SALM<T>

Models & Types · AiDotNet.SpeechRecognition.LLMIntegrated

SALM (Chen et al., 2023): a speech-augmented language model that feeds a Fast Conformer encoder’s output through a Conformer modality adapter into a Megatron GPT with LoRA, for multitask speech recognition and translation with in-context learning.

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

Implements: ISpeechRecognizer<T>

For Beginners

SALM listens with a speech encoder, shortens what it hears to one token per 320 ms with two more Conformer layers, and places those tokens in front of a written instruction for a large language model. The language model then writes the answer, here a transcript. Words given in the instruction’s text context can steer the answer (the paper’s keyword boosting) without retraining.

How It Works

§3.1 of the paper: the audio encoder is the 110M Fast Conformer-large; the modality adapter is two Conformer layers with 4x subsampling, giving a 320 ms frame shift, projected to the LLM’s width and prepended to the text context and instruction; the LLM is NVIDIA’s 2B Megatron GPT (GPT-2B-001: 24 layers of width 2,048, SwiGLU, rotary embeddings on half of each head, LayerNorm, no biases, untied embeddings, a 256,000-piece SentencePiece vocabulary), with LoRA of rank 128 added; the LLM is frozen and the rest is trained. NeMo’s modular_speechllm release fixes the rest: the adapter is NeMo’s ConformerEncoder with depthwise-separable striding; LoRA sits on the fused query-key-value projection; the template is Q: {input}\nA: {output} without a beginning-of-sequence token and with </s> after the answer; the loss covers the answer only. §4 trains with Adam (learning rate 1e-4, weight decay 1e-3), cosine annealing after 2,000 warm-up steps and gradients clipped to 5. A model built here starts from random weights and a frozen random LLM cannot fit a transcript, so the default updates every parameter; the paper’s scope is EncoderAdapterAndLora.

Example

var architecture = new NeuralNetworkArchitecture<float>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.SpeechRecognition,
    inputSize: 160000, outputSize: 256000);
var model = new SALM<float>(architecture);
var text = model.Transcribe(waveform).Text;

Constructors

Constructor Summary
SALM(NeuralNetworkArchitecture<T> architecture, SALMOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
SALM(NeuralNetworkArchitecture<T> architecture, string modelPath, SALMOptions 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 LLM’s greedy answer (the paper’s default decoding) to the instruction, at most MaxTextLength tokens.
string DetectLanguage(Tensor<T> audio)