encoder_worker.js

/**
 * Encoder Worker - Runs preprocessor + encoder in a separate thread
 */

importScripts('https://cdn.jsdelivr.net/npm/onnxruntime-web@1.21.0/dist/ort.all.min.js');

// Configure ONNX Runtime paths
ort.env.wasm.wasmPaths = 'https://cdn.jsdelivr.net/npm/onnxruntime-web@1.21.0/dist/';

const MODEL_CACHE_NAME = 'moonshine-models-v1';

// Helper to fetch model with progress reporting and caching
async function fetchModelWithProgress(url, modelName) {
    // Try to get from cache first
    try {
        const cache = await caches.open(MODEL_CACHE_NAME);
        const cachedResponse = await cache.match(url);

        if (cachedResponse) {
            const buffer = await cachedResponse.arrayBuffer();
            self.postMessage({
                type: 'progress',
                model: modelName,
                loaded: buffer.byteLength,
                total: buffer.byteLength,
                done: true,
                cached: true
            });
            console.log(`${modelName} loaded from cache`);
            return buffer;
        }
    } catch (e) {
        console.warn('Cache API not available:', e.message);
    }

    // Fetch from network
    const response = await fetch(url);
    if (!response.ok) {
        throw new Error(`Failed to fetch ${modelName}: ${response.status}`);
    }

    const contentLength = response.headers.get('Content-Length');
    const total = contentLength ? parseInt(contentLength, 10) : 0;

    if (!response.body || !total) {
        // No streaming support or unknown size - just download
        const buffer = await response.arrayBuffer();
        self.postMessage({
            type: 'progress',
            model: modelName,
            loaded: buffer.byteLength,
            total: buffer.byteLength,
            done: true
        });
        // Cache the response
        try {
            const cache = await caches.open(MODEL_CACHE_NAME);
            await cache.put(url, new Response(buffer.slice(0)));
        } catch (e) {
            console.warn('Failed to cache model:', e.message);
        }
        return buffer;
    }

    const reader = response.body.getReader();
    const chunks = [];
    let loaded = 0;

    while (true) {
        const { done, value } = await reader.read();
        if (done) break;

        chunks.push(value);
        loaded += value.length;

        self.postMessage({
            type: 'progress',
            model: modelName,
            loaded,
            total,
            done: false
        });
    }

    self.postMessage({
        type: 'progress',
        model: modelName,
        loaded: total,
        total,
        done: true
    });

    // Combine chunks into single ArrayBuffer
    const result = new Uint8Array(loaded);
    let offset = 0;
    for (const chunk of chunks) {
        result.set(chunk, offset);
        offset += chunk.length;
    }

    // Cache the result
    try {
        const cache = await caches.open(MODEL_CACHE_NAME);
        await cache.put(url, new Response(result.slice(0)));
        console.log(`${modelName} cached`);
    } catch (e) {
        console.warn('Failed to cache model:', e.message);
    }

    return result.buffer;
}

// Model config
let cfg = null;
let preprocessor = null;
let encoder = null;
let tailLatency = 0;

// Preprocessor state
let prepSession = null;
let prepDim = 0;
let prepC1 = 0;
let prepStateC1 = null;
let prepStateC2 = null;

// Encoder state
let encSession = null;
let encDim = 0;
let encNPast = 0;
let encNFuture = 0;
let encEncoderDepth = 0;
let encContextSize = 0;
let encInputBuffer = [];
let encTotalInputFrames = 0;
let encTotalOutputFrames = 0;

function resetPreprocessor() {
    if (prepStateC1) prepStateC1.fill(0);
    if (prepStateC2) prepStateC2.fill(0);
}

function resetEncoder() {
    encInputBuffer = [];
    encTotalInputFrames = 0;
    encTotalOutputFrames = 0;
}

async function processPreprocessor(audioChunk) {
    const feeds = {
        'audio_chunk': new ort.Tensor('float32', audioChunk, [1, audioChunk.length]),
        'state_c1': new ort.Tensor('float32', prepStateC1, [1, 4, prepDim]),
        'state_c2': new ort.Tensor('float32', prepStateC2, [1, 4, prepC1])
    };

    const results = await prepSession.run(feeds);

    // Update states
    prepStateC1.set(results.new_state_c1.data);
    prepStateC2.set(results.new_state_c2.data);

    return {
        data: results.features.data,
        dims: results.features.dims
    };
}

async function processEncoder(melData, melDims, flush = true) {
    const newFrames = melDims[1];

    // Append new frames to buffer
    for (let f = 0; f < newFrames; f++) {
        const frame = new Float32Array(encDim);
        for (let d = 0; d < encDim; d++) {
            frame[d] = melData[f * encDim + d];
        }
        encInputBuffer.push(frame);
    }

    encTotalInputFrames += newFrames;

    // Calculate output range
    const canOutput = flush
        ? encTotalInputFrames
        : Math.max(0, encTotalInputFrames - tailLatency);

    const outputFrom = flush
        ? Math.max(0, encTotalOutputFrames - tailLatency)
        : encTotalOutputFrames;

    const newOutputCount = canOutput - outputFrom;

    if (newOutputCount <= 0) {
        return { data: new Float32Array(0), dims: [1, 0, encDim] };
    }

    // Prepare input buffer tensor
    const bufferFrames = encInputBuffer.length;
    const bufferData = new Float32Array(bufferFrames * encDim);
    for (let f = 0; f < bufferFrames; f++) {
        bufferData.set(encInputBuffer[f], f * encDim);
    }

    const feeds = {
        'input': new ort.Tensor('float32', bufferData, [1, bufferFrames, encDim])
    };

    const results = await encSession.run(feeds);
    const fullOutput = results.output;

    // Calculate which frames to return
    const bufStartFrame = encTotalInputFrames - bufferFrames;
    const outputStart = outputFrom - bufStartFrame;

    // Extract the subset of output
    const resultData = new Float32Array(newOutputCount * encDim);
    for (let f = 0; f < newOutputCount; f++) {
        for (let d = 0; d < encDim; d++) {
            resultData[f * encDim + d] = fullOutput.data[(outputStart + f) * encDim + d];
        }
    }

    // Trim input buffer to context size
    if (encInputBuffer.length > encContextSize) {
        encInputBuffer = encInputBuffer.slice(-encContextSize);
    }

    encTotalOutputFrames = canOutput;
    return { data: resultData, dims: [1, newOutputCount, encDim] };
}

// Message queue for sequential processing
const messageQueue = [];
let isProcessing = false;

async function processMessage(e) {
    const { type, data } = e.data;

    switch (type) {
        case 'init': {
            try {
                cfg = data.cfg;
                const onnxUrl = data.onnxUrl;
                const modelName = data.modelName;
                const backend = data.backend || 'wasm';
                const dtype = 'fp32';

                // Check WebGPU availability
                if (backend === 'webgpu') {
                    if (typeof navigator !== 'undefined' && navigator.gpu) {
                        console.log('WebGPU navigator.gpu is available');
                        const adapter = await navigator.gpu.requestAdapter();
                        if (adapter) {
                            console.log('WebGPU adapter obtained:', adapter);
                        } else {
                            throw new Error('WebGPU adapter not available');
                        }
                    } else {
                        throw new Error('WebGPU not supported (navigator.gpu is undefined)');
                    }
                }

                const sessionOptions = { executionProviders: [backend] };
                console.log(`Creating sessions with backend: ${backend}`);

                tailLatency = cfg.n_future * cfg.encoder_depth;

                // Initialize preprocessor
                const prepUrl = `${onnxUrl}/preprocessor_streaming_${modelName}_${dtype}.onnx`;
                self.postMessage({ type: 'status', message: 'Loading preprocessor...' });
                self.postMessage({ type: 'model_start', model: 'Preprocessor' });
                const prepBuffer = await fetchModelWithProgress(prepUrl, 'Preprocessor');
                prepSession = await ort.InferenceSession.create(prepBuffer, sessionOptions);
                self.postMessage({ type: 'model_done', model: 'Preprocessor' });

                prepDim = cfg.dim;
                prepC1 = 2 * cfg.dim;
                prepStateC1 = new Float32Array(4 * cfg.dim);
                prepStateC2 = new Float32Array(4 * prepC1);

                // Initialize encoder
                const encUrl = `${onnxUrl}/encoder_${modelName}_${dtype}.onnx`;
                self.postMessage({ type: 'status', message: 'Loading encoder...' });
                self.postMessage({ type: 'model_start', model: 'Encoder' });
                const encBuffer = await fetchModelWithProgress(encUrl, 'Encoder');
                encSession = await ort.InferenceSession.create(encBuffer, sessionOptions);
                self.postMessage({ type: 'model_done', model: 'Encoder' });

                encDim = cfg.dim;
                encNPast = cfg.n_past;
                encNFuture = cfg.n_future;
                encEncoderDepth = cfg.encoder_depth;
                encContextSize = cfg.encoder_depth * (cfg.n_past + cfg.n_future);

                self.postMessage({ type: 'ready', backend: backend });
            } catch (err) {
                self.postMessage({ type: 'error', message: err.message });
            }
            break;
        }

        case 'segment_start': {
            resetPreprocessor();
            resetEncoder();
            self.postMessage({
                type: 'segment_start',
                segmentId: data.segmentId
            });
            break;
        }

        case 'segment_end': {
            self.postMessage({
                type: 'segment_end',
                segmentId: data.segmentId
            });
            break;
        }

        case 'audio': {
            try {
                // Process through preprocessor
                const mel = await processPreprocessor(new Float32Array(data.audio));

                const audioMs = (data.audio.length / 16000 * 1000).toFixed(0);
                console.log(`Audio ${data.audio.length} samples (${audioMs}ms) → Mel ${mel.dims[1]} frames`);

                // Process through encoder with flush=true
                const enc = await processEncoder(mel.data, mel.dims, true);

                console.log(`Mel ${mel.dims[1]} frames → Encoder ${enc.dims[1]} frames (accumulated: ${encTotalOutputFrames})`);

                if (enc.dims[1] > 0) {
                    self.postMessage({
                        type: 'features',
                        segmentId: data.segmentId,
                        features: enc.data,
                        dims: enc.dims
                    }, [enc.data.buffer]);  // Transfer ownership
                }
            } catch (err) {
                console.error('Encoder error:', err);
            }
            break;
        }
    }
}

async function processQueue() {
    if (isProcessing) return;
    isProcessing = true;

    while (messageQueue.length > 0) {
        const msg = messageQueue.shift();
        await processMessage(msg);
    }

    isProcessing = false;
}

self.onmessage = function(e) {
    messageQueue.push(e);
    processQueue();
};