Voices

Early Beta — The Web SDK is in early beta. APIs may change between releases.

Overview

Browser text-to-speech uses Piper VITS voices in ONNX format, rendered by the sherpa-onnx WASM build in @runanywhere/web-onnx. A voice is a catalog model with category MODEL_CATEGORY_SPEECH_SYNTHESIS; there is no separate voice loader.

Registering a voice

Piper voices ship as .tar.gz bundles, so they register through registerModelArchive.

import { RunAnywhere } from '@runanywhere/web'
import {
  InferenceFramework,
  ModelArtifactType,
  ModelCategory,
  ModelFormat,
} from '@runanywhere/proto-ts/model_types'

RunAnywhere.registerModelArchive(
  'https://huggingface.co/runanywhere/vits-piper-en_US-lessac-medium/resolve/main/vits-piper-en_US-lessac-medium.tar.gz',
  'Piper TTS (US English - Medium)',
  InferenceFramework.INFERENCE_FRAMEWORK_SHERPA,
  ModelArtifactType.MODEL_ARTIFACT_TYPE_TAR_GZ_ARCHIVE,
  {
    id: 'vits-piper-en_US-lessac-medium',
    description: 'Piper VITS text-to-speech, medium quality.',
    modality: ModelCategory.MODEL_CATEGORY_SPEECH_SYNTHESIS,
    format: ModelFormat.MODEL_FORMAT_ONNX,
    downloadSizeBytes: 67_389_394,
    memoryRequirement: 90_000_000,
  }
)

Voices in the example app

Model id	Language	Download size
`vits-piper-en_US-lessac-medium`	English (US)	~67 MB
`vits-piper-en_GB-alba-medium`	English (UK)	~67 MB

Piper publishes hundreds of voices across many languages. Browse the catalog at Piper Samples and register any of them the same way. Whatever you point at must be reachable from the browser, so it needs permissive CORS headers.

Loading and switching

Loading a speech-synthesis model replaces whichever voice was loaded for that category.

import { RunAnywhere, ModelCategory } from '@runanywhere/web'

await RunAnywhere.downloadModel('vits-piper-en_GB-alba-medium')

const result = await RunAnywhere.loadModel({
  modelId: 'vits-piper-en_GB-alba-medium',
  category: ModelCategory.MODEL_CATEGORY_SPEECH_SYNTHESIS,
  forceReload: false,
  validateAvailability: true,
})

if (!result?.success) {
  console.error(result?.errorMessage ?? 'Load failed')
}

await RunAnywhere.speak('Hello from the British English voice.')

To free it:

await RunAnywhere.unloadModel({
  modelId: 'vits-piper-en_GB-alba-medium',
  category: ModelCategory.MODEL_CATEGORY_SPEECH_SYNTHESIS,
  unloadAll: false,
})

Enumerating voices inside a model

A multi-speaker Piper model exposes several voices. RunAnywhere.tts.listLoadedVoices() returns what the currently loaded engine can render.

const voices = RunAnywhere.tts.listLoadedVoices()
for (const voice of voices) {
  console.log(voice.id, voice.displayName, voice.languageCode, voice.sampleRate)
}

interface TTSVoiceInfo {
  id: string
  displayName: string
  languageCode: string
  gender: TTSVoiceGender
  description: string
  isNeural: boolean
  isSystem: boolean
  sampleRate: number
  supportedStyles: string[]
}

The list is empty when no TTS model is loaded or the WASM build lacks the enumeration export.

Selecting a voice or speaker

// By engine voice id
await RunAnywhere.speak('Hello!', { voiceId: 'en_US-lessac-medium' })

// By speaker index within a multi-speaker model
const speaker0 = await RunAnywhere.synthesize('Hello!', { speakerId: 0 })
const speaker1 = await RunAnywhere.synthesize('Hello!', { speakerId: 1 })

A voice id is not a model id. The model id names the downloaded bundle; the voice id names a voice inside it. Passing a model id as voiceId will not select anything.

Checking what is loaded

import { RunAnywhere, ModelCategory, SDKComponent } from '@runanywhere/web'

const model = RunAnywhere.modelInfoForCategory(ModelCategory.MODEL_CATEGORY_SPEECH_SYNTHESIS)
console.log(model?.id)

const snapshot = RunAnywhere.componentLifecycleSnapshot(SDKComponent.SDK_COMPONENT_TTS)
console.log(snapshot?.state)

Handle API

RunAnywhere.tts also exposes a handle-based surface for callers that want to own the component lifecycle: create(), loadVoice(handle, path, id?, name?), isLoaded(handle), listVoices(handle), synthesize(handle, text, options?), synthesizeStream(handle, text, options?), stop(handle), unload(handle), and destroy(handle). Prefer the flat RunAnywhere.synthesize / speak path.

Synthesize

Text-to-speech synthesis

Streaming TTS

Chunked synthesis and playback

Voice agent

Full voice turn

Getting Started

Swift SDK

Kotlin SDK

React Native SDK

Flutter SDK

Web SDK

Vibe Coding

Overview

Registering a voice

Voices in the example app

Loading and switching

Enumerating voices inside a model

Selecting a voice or speaker

Checking what is loaded

Handle API

Synthesize

Streaming TTS

Voice agent

​Overview

​Registering a voice

​Voices in the example app

​Loading and switching

​Enumerating voices inside a model

​Selecting a voice or speaker

​Checking what is loaded

​Handle API

​Related

Synthesize

Streaming TTS

Voice agent

Overview

Registering a voice

Voices in the example app

Loading and switching

Enumerating voices inside a model

Selecting a voice or speaker

Checking what is loaded

Handle API

Related