Download engine.mjs from dunli5107/SmollerTTS: direct link, hf CLI and curl.
- Browser
- Download file 3.72 kB
-
https://huggingface.co/dunli5107/SmollerTTS/resolve/main/engine.mjs
- Command line
-
hf download hf://dunli5107/SmollerTTS/engine.mjs
-
curl -L -o engine.mjs https://huggingface.co/dunli5107/SmollerTTS/resolve/main/engine.mjs
3.72 kB
| // Preserve the accepted unfused arithmetic. Graph fusion failed the v16 parity audit. | |
| import * as ort from './ort/ort.wasm.min.mjs'; | |
| import {validateControls, adjustDurations, adjustProsody} from './controls-v19.mjs'; | |
| import {applyPunctuation} from './punctuation.mjs'; | |
| export async function loadEngine(models, wasmBinary) { | |
| ort.env.wasm.numThreads = 1; | |
| ort.env.wasm.wasmPaths = new URL('./ort/', import.meta.url).href; | |
| if (wasmBinary) ort.env.wasm.wasmBinary = wasmBinary; | |
| const sessions = {}; | |
| try { | |
| // Downloads are parallel, session construction is serial: one shared WASM runtime. | |
| for (const name of ['encoder', 'frames', 'vocoder']) { | |
| sessions[name] = await ort.InferenceSession.create(models[name], | |
| {executionProviders: ['wasm'], graphOptimizationLevel: 'disabled'}); | |
| } | |
| } catch (error) { | |
| for (const session of Object.values(sessions)) await session.release(); | |
| throw error; | |
| } | |
| const integer = (data, dims) => new ort.Tensor('int64', BigInt64Array.from(data, BigInt), dims); | |
| return { | |
| async run({tokens, stress, durations, punctuation = [], controls = {}}) { | |
| controls = validateControls(controls); | |
| const owned = [], keep = tensor => (owned.push(tensor), tensor); | |
| const start = performance.now(), timings = {}; | |
| try { | |
| const encoded = await sessions.encoder.run({tokens: keep(integer(tokens, [1, tokens.length])), stress: keep(integer(stress, [1, tokens.length]))}); | |
| timings.encoder = performance.now() - start; | |
| Object.values(encoded).forEach(keep); | |
| durations ??= Array.from(encoded.duration.data, (d, i) => { | |
| const n = Math.round(Math.expm1(Math.max(0, Math.min(Math.log(101), d)))); | |
| return tokens[i] === 1 ? Math.max(0, Math.min(60, n)) : Math.max(1, Math.min(100, n)); | |
| }); | |
| const shaped = applyPunctuation(tokens, stress, durations, adjustProsody(encoded.phone_prosody.data, controls), punctuation, controls); | |
| durations = adjustDurations(shaped.durations, tokens, controls.speed); | |
| const indices = [], position = []; | |
| for (let i = 0; i < durations.length; i++) for (let j = 0; j < durations[i]; j++) { | |
| indices.push(i); position.push((j + .5) / durations[i]); | |
| } | |
| if (indices.length < 2 || indices.length > 6000) throw new Error('Please use a shorter sentence.'); | |
| const changed = shaped.prosody; | |
| const phoneProsody = changed === encoded.phone_prosody.data ? encoded.phone_prosody : | |
| keep(new ort.Tensor('float32', changed, encoded.phone_prosody.dims)); | |
| const frameStart = performance.now(); | |
| const result = await sessions.frames.run({encoded: encoded.encoded, phone_prosody: phoneProsody, | |
| indices: keep(integer(indices, [1, indices.length])), position: keep(new ort.Tensor('float32', Float32Array.from(position), [1, indices.length]))}); | |
| Object.values(result).forEach(keep); timings.frames = performance.now() - frameStart; | |
| const vocoderStart = performance.now(); | |
| const audio = await sessions.vocoder.run({mel: result.mel}); Object.values(audio).forEach(keep); | |
| timings.vocoder = performance.now() - vocoderStart; | |
| // Copy once out of WASM; avoid expanding PCM into boxed JavaScript numbers. | |
| const wave = new Float32Array(audio.waveform.data); | |
| if (wave.some(value => !Number.isFinite(value))) throw new Error('Speech generation produced invalid audio.'); | |
| return {wave, milliseconds: performance.now() - start, timings, durations, controls}; | |
| } finally { for (const tensor of owned) tensor.dispose(); } | |
| }, | |
| async close() {for (const session of Object.values(sessions)) await session.release();} | |
| }; | |
| } | |