@met4citizen/headtts v0.2.0

  HeadTTS

HeadTTS is a free JavaScript text-to-speech (TTS) solution that provides timestamps and Oculus visemes for lip-sync, in addition to audio output (WAV/PCM). It uses Kokoro neural model and voices, and inference can run entirely in the browser (via WebGPU or WASM), or alternatively on a Node.js WebSocket/RESTful server (CPU-based).

• Pros: Free. Doesn't require a server in in-browser mode. WebGPU support. Uses neural voices with a StyleTTS 2 model. Great for lip-sync use cases and fully compatible with the TalkingHead. MIT licensed, doesn't use eSpeak NG or any other GPL-licensed module.

• Cons: WebGPU is only supported by default in Chrome and Edge desktop browsers. Takes time to load (the first time) and uses a lot of memory. WebGPU support in onnxruntime-node is still experimental and not released, so server-side inference is relatively slow. English is currently the only supported language.

If you're using a Chrome or Edge desktop browser, check out the In-browser Demo!

The project uses websockets/ws (MIT License), hugginface/transformers.js (Apache 2.0 License) and onnx-community/Kokoro-82M-v1.0-ONNX-timestamped (Apache 2.0 License) as runtime dependencies. For information on language modules and dictionaries, see Appendix B. Using jest for testing.

You can find the list of supported English voices and voice samples here.

In-browser Module: headtts.mjs

The HeadTTS JavaScript module enables in-browser text-to-speech using Module Web Workers and WebGPU/WASM inference. Alternatively, it can connect to and use the HeadTTS Node.js WebSocket/RESTful server.

Create a new HeadTTS class instance:

import { HeadTTS } from "./modules/headtts.mjs";

const headtts = new HeadTTS({
  endpoints: ["ws://127.0.0.1:8882", "webgpu"], // Endpoints in order of priority
  languages: ['en-us'], // Language modules to pre-load (in-browser)
  voices: ["af_bella", "am_fenrir"] // Voices to pre-load (in-browser)
});

Note: Model related options apply only to in-browser inference. If inference is performed on a server, server-specific settings will apply instead.

Connect to the first supported/available endpoint:

try {
  await headtts.connect();
} catch(error) {
  console.error(error);
}

Make an onmessage event handler to handle response messages. In this example, we use TalkingHead instance head to play the incoming audio and lip-sync data:

// Speak and lipsync
headtts.onmessage = (message) => {
  if ( message.type === "audio" ) {
    try {
      head.speakAudio( message.data, {}, (word) => {
        console.log(word);
      });
    } catch(error) {
      console.log(error);
    }
  } else if ( message.type === "error" ) {
    console.error("Received error message, error=", message.data.error);
  }
}

Setup the voice only when you need it to change:

headtts.setup({
  voice: "af_bella",
  language: "en-us",
  speed: 1,
  audioEncoding: "wav"
});

Synthesize speech:

headtts.synthesize({
  input: "Test sentence."
});

The above approach relies on onmessage event handler to receive and handle response messages and it is the recommended approach for real-time use cases. An alternative approach is to await for all the related audio messages:

try {
  const messages = await headtts.synthesize({
    input: "Some long text..."
  });
  console.log(messages); // [{type: 'audio', data: {…}, ref: 1}, {…}, ...]
} catch(error) {
  console.error(error);
}

The input parameter can be a string or, alternatively, an array of strings or inputs items.

TODO: Add support for audio type.

An example using an array of input items:

{
  type: "synthesize",
  id: 14, // Unique request identifier.
  data: {
    input: [
      "There were ",
      { type: "speech", value: "over two hundred ", subtitles: ">200 " },
      "items of",
      { type: "phonetic", value: "mˈɜɹʧəndˌIz ", subtitles: "merchandise " },
      "on sale."
    ]
  }
}

NodeJS WebSocket/RESTful Server: headtts-node.mjs

Install using NPM (requires Node.js v20+):

npm install @met4citizen/headtts

Start the server:

npm start

An example:

node ./modules/headtts-node.mjs --trace 16

Appendix A: Server API reference

WebSocket API

Every WebSocket request must have a unique identifier, id. The server uses a Web Worker thread pool, and because work is done in parallel, the order of responses may vary. Therefore, each response includes a ref property that identifies the original request, allowing the order to be restored if necessary. The JS client class handles this automatically.

Request: setup

{
  type: "setup",
  id: 12, // Unique request identifier.
  data: {
    voice: "af_bella", // Voice name (optional)
    language: "en-us", // Language (optional)
    speed: 1, // Speed (optional)
    audioEncoding: 'wav' // "wav" or "pcm" (PCM 16bit LE) (optional)
  }
}

Request: synthesize

{
  type: "synthesize",
  id: 13, // Unique request identifier.
  data: {
    input: "This is an example." // String or array of input items
  }
}

The response message for synthesize request is either error or audio.

Response: error

{
  type: "error",
  ref: 13, // Original request id
  data: {
    error: "Error loading voice 'af_bella'."
  }
}

Response: audio

Returns an audio object metadata that can be passed on the TalkingHead speakAudio method once the audio content itself has been added.

{
  type: "audio",
  ref: 13,
  data: {
    words: ["This ","is ","an ","example."],
    wtimes: [443, 678, 780, 864],
    wdurations: [192 ,91 ,52 ,868],
    visemes: ["TH", "I", "SS", "I", "SS", "aa", "nn", "SS", "aa", "PP", "PP", "E", "DD"],
    vtimes: [443, 494, 550, 678, 729, 780, 801, 989, 1126, 1175, 1225, 1275, 1354],
    vdurations: [61, 66, 85, 61, 40, 31, 31, 69, 59, 60, 60, 89, 378],
    audioEncoding: "wav"
  }
}

The actual audio content will be delivered after this message as binary data (see the next response message).

Response: Binary (ArrayBuffer)

Binary data as an ArrayBuffer related to the previous audio message. Depending on the set audio encoding, either a WAV file (wav) or a chunk of raw PCM 16bit LE samples (pcm).

RESTful API

RESTful server API is a more simple alternative for WebSocket API.

POST /v1/synthesize

OK response:

Error response:

Appendix B: Language modules and dictionaries

American English, en-us

The American English language module is based on the CMU Pronunciation Dictionary from Carnegie Mellon University, containing over 134,000 words and their pronunciations. The original dataset is provided under a simplified BSD license, allowing free use for any research or commercial purpose.

In the Kokoro TTS model, the American English language data was trained using the Misaki G2P engine (en). Therefore, the original ARPAbet phonemes in the CMU dictionary have been converted to IPA and then to Misaki-compatible phonemes by applying the following mapping:

• ɚ → ɜ, ɹ , ˈɝ → ˈɜ, ɹ , ˌɝ → ˌɜ, ɹ
• tʃ → ʧ , dʒ → ʤ
• eɪ → A , ˈeɪ → ˈA , ˌeɪ → ˌA
• aɪ → I , ˈaɪ → ˈI , ˌaɪ → ˌI
• aʊ → W , ˈaʊ → ˈW , ˌaʊ → ˌW
• ɔɪ → Y , ˈɔɪ → ˈY , ˌɔɪ → ˌY
• oʊ → O , ˈoʊ → ˈO , ˌoʊ → ˌO
• əʊ → Q , ˈəʊ → ˈQ , ˌəʊ → ˌQ

Note: During the dataset conversion, some vowels were reduced to reflect casual speech since HeadTTS is primarily designed for conversational use.

The final dictionary is a plain text file with around 125,000 lines (2,8MB). Lines starting with ;;; are comments. Each other line represents one word and its pronunciations. The word and its different possible pronunciations are separated by a tab character \t. An example entry:

MERCHANDISE	mˈɜɹʧəndˌIz

Out-of-dictionary (OOD) words are converted using a rule-based algorithm based on NRL Report 7948, Automatic Translation of English Text to Phonetics by Means of Letter-to-Sound Rules (Elovitz et al., 1976). The report is available here.

Finnish, fi

!IMPORTANT
As of now, Finnish language is not supported by the Kokoro model. You can use the fi language code with the English voices, but the pronunciation will sound rather weird.

The phonemization of the Finnish language module is done by an in-built algorithm. The algorithm doesn't require a pronunciation dictionary, but it uses a compound word dictionary to get the secondary stress marks right for compound words.

The dictionary used for compound words is based on The Dictionary of Contemporary Finnish maintained by the Institute for the Languages of Finland. The original dataset contains more than 100,000 entries and is open-sourced under the CC BY 4.0 license.

The pre-processed compound word dictionary is a plain text file with around 50,000 entries in 10,000 lines (~350kB). Lines starting with ;;; are comments. Each other line represents the first part of a compound word and the first four letters of all possible next words, all separated by a tab character \t. An example entry:

ALUMIINI	FOLI	KATT	OKSI	PAPE	SEOS	VENE	VUOK

Appendix C: Latency

In-browser TTS on WebGPU is 3x times faster than real-time and approximately 10x faster than WASM. CPU inference on a Node.js server performs surprisingly well, but increasing the thread pool size worsens performance, so we need to wait for WebGPU support. Quantization makes no significant difference, so I recommend using 32-bit floating point precision (fp32) for the best audio quality unless memory consumption becomes a concern.

Unofficial latency results using my own latency test app:

[1] Finish latency: Total time from sending text input to receiving the full audio.

[2] First byte/part/sentence latency: Time from sending the text input to receiving the first playable byte/part/sentence of audio. Note: This measure is not comparable across all models, since some solutions use streaming, some not.

[3] Real-time factor = Time to generate full audio / Duration of the full audio. If RTF < 1, synthesis is faster than real-time (i.e., good).

Test setup: Macbook Air M2 laptop, 8 cores, 16GB memory, macOS Sequoia 15.3.2, Metal2 GPU 10 cores, 300/50 Mbit/s internet connection. The latest Google Chrome/Edge desktop browsers.

All test cases use WAV or raw PCM 16bit LE format and the "List 1" of the Harvard Sentences:

The birch canoe slid on the smooth planks.
Glue the sheet to the dark blue background.
It's easy to tell the depth of a well.
These days a chicken leg is a rare dish.
Rice is often served in round bowls.
The juice of lemons makes fine punch.
The box was thrown beside the parked truck.
The hogs were fed chopped corn and garbage.
Four hours of steady work faced us.
A large size in stockings is hard to sell.