import * as FileSystem from 'expo-file-system'; import { InferenceSession, Tensor } from "onnxruntime-react-native"; import { tokenizer } from "./tokenizer"; import { load_voice_data, KokoroVoice } from "./voices"; import floatArrayToWAV from "./wav"; import { OpenPhonemizer } from 'expo-open-phonemizer'; import { Asset } from 'expo-asset'; const SAMPLE_RATE = 24000; const STYLE_DIM = 256; const MAX_PHONEME_LENGTH = 510; export class Kokoro { phonemizer: OpenPhonemizer; session: InferenceSession; constructor(session: InferenceSession, phonemizer: OpenPhonemizer) { this.session = session; this.phonemizer = phonemizer; } static async load(): Promise { const asset = Asset.fromModule(require('./kokoro-quantized.onnx')); if (!asset.downloaded) { console.log("Downloading Kokoro model..."); await asset.downloadAsync(); } const modelPath = asset.localUri ?? asset.uri; const options = { graphOptimizationLevel: 'all', enableCpuMemArena: true, enableMemPattern: true, executionMode: 'sequential' } as InferenceSession.SessionOptions; const session = await InferenceSession.create( modelPath, options ); const phonemizer = await OpenPhonemizer.load(); return new Kokoro(session, phonemizer); } async generate(text: string, voice: KokoroVoice, outputPath: string): Promise { const phonemes = await this.phonemizer.phonemize(text, true); const tokens = tokenizer.encode(phonemes); const n_tokens = Math.min(Math.max(tokens.length - 2, 0), MAX_PHONEME_LENGTH - 1); const offset = n_tokens * STYLE_DIM; const voice_data = await load_voice_data(voice); const style_data = voice_data.slice(offset, offset + STYLE_DIM); const inputs = { input_ids: new Tensor('int64', tokens, [1, tokens.length]), style: new Tensor('float32', new Float32Array(style_data), [1, STYLE_DIM]), speed: new Tensor('float32', new Float32Array([1.0]), [1]) } const outputs = await this.session.run(inputs); if (!outputs || !outputs['waveform'] || !outputs['waveform'].data) { throw new Error('Invalid output from model inference'); } const data = outputs['waveform'].data as Float32Array; const wav = floatArrayToWAV(data, SAMPLE_RATE); await FileSystem.writeAsStringAsync( outputPath, wav, { encoding: FileSystem.EncodingType.Base64 } ); } }