Feature Library

Offline narrator

<0.01 CPU ms
A still of the Offline narrator demo
Still
The live demo starts when it’s in view.Edit the script, insert marks, pick a voice · Play script or Ctrl Enter on a line · Another take, then keep one · Esc stops

Narration from a script with no cloud: lines with pauses and tone marks, a pronunciation list and {values}, parsed into pieces that keep their place in the…

What it costs a frame

<0.01 ms
GPU
not used
CPU, median
<0.01 ms
CPU, 95th pct
<0.01 ms
Draw calls
0
Weight
Light

Measured 11 November 2026 on 13th Gen Intel(R) Core(TM) i9-13900H, Chrome 157, 1920×1080 at 1×. 600 frames × 3 rounds, median kept. What ran: Measured with the demo exercising itself: the five-line script plays on a loop through the silent timing engine (the bench has no click, so the system voice may not speak), with the word pill moving in the editor and the reading view, the ribbon pulsing on each word, the pause rings and the timeline playhead. Parsing and highlighting run on each edit, not each frame. Phones: tiers are being measured next.

On every machine it was measured on
MachineGPU msCPU msCPU p95
unknown GPU · 13th Gen Intel(R) Core(TM) i9-13900Hreference · measuring toolnot timed<0.01<0.01
apple · Apple A19 Proin a browser tab · 16 November 2026not timed<0.01<0.01
Intel Arc 140V (Xe2-LPG) · Intel Core Ultra 7 258Vin a browser tab · 15 November 2026not timed<0.01<0.01

Bars against one 16.7 ms frame. A browser-tab run is timed in a visible tab without the tool’s scripted input, so read it as close, not exact.

What it needs

Renders with
DOM and CSSPlain DOM and CSS: every current browser.
Your browser
Runs hereNo WebGPU needed
Depends on
NothingNo other packages

What it’s for

  1. Narration for a presentation, a tour or a walkthrough
  2. Voice-over drafts before a real recording
  3. Reading a script aloud while writing it
  4. Kiosks and exhibits that must run without a network
  5. Spoken prompts in a game or a tutorial
  6. Clips made once on a build machine and played anywhere

Pitfalls, in brief

  1. A copied voice needs the speaker’s consent, in writing, for the use you make of it.
  2. Not every system voice is offline.
  3. The voice list fills after the page loads (Chrome fires voiceschanged; some browsers don’t): voices() waits for it up to voiceWait.
  4. Some voices give no word boundary events (several online ones, some on Android).
  5. Chrome loses an utterance’s events if nothing holds a reference to it, and some voices stop after about 15 seconds of one utterance.
5 more in brief
  1. A live engine can’t hand back its audio: the system voice speaks straight to the speakers, so it can’t be cached, mixed under music or saved.
  2. The clip key has the engine’s version, the voice, the recording’s stamp, the settings that change the sound and the spoken words with their marks.
  3. Pronunciations match whole words, case-sensitively, longest first, and are applied after {values} are filled.
  4. Kokoro takes about as long as the line lasts on a laptop CPU; Chatterbox about 1.4 times real time on a laptop GPU and several gigabytes on disk.
  5. The voice service names Node’s modules as inputs (fs, path, spawn): pass them in.
10 pitfalls in full, the hook-up guide, the module source and its tests come with the part.

Where it comes from

From My other apps. Fully custom code.

Web Speech API (W3C Community Group): speechSynthesis, its voices (localService marks the ones on the device) and the word boundary events

Kokoro-82M (hexgrad), Apache 2.0: the CPU engine; run through kokoro-js (Apache 2.0) and Transformers.js. Not included: downloaded by you, once

Chatterbox (Resemble AI), MIT: the GPU engine and its voice copying, run by chatterbox_worker.py. Not included: installed by you, once

pyloudnorm (Christian Steinmetz), MIT: the worker levels every clip to −18 LUFS (ITU-R BS.1770 loudness)

FNV-1a hash (Fowler, Noll, Vo), public domain: the 64-bit clip names, written here from the published constants

George Marsaglia, “Xorshift RNGs” (2003): the small generator behind each take’s pitch and pace

WAI-ARIA Authoring Practices Guide (W3C): the tabs, radio group and toolbar patterns in the demo

text to speechnarrationvoiceweb speechspeech synthesisofflinekokorochatterboxvoice cloningpronunciationcachequeuewavnodeaccessibilitykeyboarddomno dependencies

Near it in the frame