Text to Speech

What is the Web Speech API, and how does synthesis work?

The Web Speech API is a W3C Community Group specification with two halves: SpeechRecognition (microphone → text) and SpeechSynthesis (text → audio). This tool uses the synthesis half, exposed in every modern browser as window.speechSynthesis. The API is deliberately thin: you create a SpeechSynthesisUtterance carrying text plus optional voice, rate, pitch and volume, and the browser forwards it to the operating system’s native speech engine. On macOS and iOS that is Apple’s speech stack (the same voices behind Siri and Spoken Content); on Windows it is SAPI 5 or the newer OneCore neural voices; on Android it is Google Text-to-Speech; on Linux it is typically eSpeak-NG or Festival. The webpage never synthesizes anything itself — it is a remote control for software already on your machine.

Voices: where they come from and why they differ per device

Calling speechSynthesis.getVoices() returns a list of SpeechSynthesisVoice objects, each with a name, a BCP-47 language tag (like en-US or ja-JP), and flags for local vs. remote service. Two subtleties matter. First, browsers populate this list asynchronously: on first page load the list is usually empty and fills in when the voiceschanged event fires, so a correct implementation must listen for that event rather than reading voices once during render. Second, the set is a property of your OS installation — installing a new voice in system settings (Windows’ “Add voices”, macOS’s downloadable Siri voices, Android’s TTS settings) makes it appear here automatically. Quality varies enormously: decade-old formant voices sound robotic, while modern neural voices (Microsoft’s OneCore, Apple’s Siri voices) are near-human. This tool groups voices by language and prefers voices reported as local. If none are available, it offers network voices with a notice that their service may send text to its servers. The number of voices depends on your browser and operating system.

Rate, pitch, and what they actually change

The utterance’s rate (0.1–10, practical range 0.5–2) scales the phoneme durations the engine produces — it is true time-stretching of the synthesis, not playback resampling, so pitch does not rise when you speed up. pitch (0–2) shifts the fundamental frequency of the generated voice, which reads as a deeper or higher voice rather than chipmunk-style resampling. volume (0–1) is a simple gain. Support is voice-dependent: some compact embedded voices ignore pitch entirely, and a few cap the rate range. The sliders here update the next utterance immediately; changing them mid-speech applies to the following sentence chunk.

Why long text is split into sentences

Chromium has a long-standing, still-unfixed bug: a single utterance longer than roughly 15 seconds of speech may stop without firing an end event, leaving the queue silently dead. The robust workaround — used by every serious browser-TTS implementation — is to split text into sentence-sized chunks (here, capped around 220 characters) and speak them as a chained queue, starting the next chunk in the previous chunk’s onend handler. This also creates natural pause points and makes the Stop button responsive: cancelling clears the queue instead of waiting out one giant utterance. Word highlighting rides on the boundary event, which reports the character offset of the word being spoken; the UI maps that offset back to word boundaries in the original text. Note that some voices never fire boundary events — highlighting then simply stays off rather than breaking.

The download question, answered honestly

The most-requested TTS feature is “download as MP3”, and the honest answer is that the Web Speech API cannot do it. Synthesis output goes directly to the OS audio pipeline; the specification exposes no MediaStream, no AudioBuffer, and no file handle, so there is nothing for MediaRecorder or the Web Audio API to capture. This is a deliberate platform limitation, not a missing feature of this page. Services that offer downloads synthesize on their own servers — which inherently means uploading your text. Anyone who needs an audio file has three honest paths: use a server-side TTS service and accept the upload, record system audio locally with a tool like Audacity or OBS, or run an offline neural TTS model (Piper, Coqui, Sherpa-ONNX) entirely on their own machine. This page chooses transparency over a fake download button.

Privacy model and practical uses

This page hands text to the browser’s voice service. Local voices synthesize on your device. When no local voice is available, network voices may send text to their servers and the tool displays that limitation. A voice name does not prove it works offline; check the browser’s local-service flag and install a local voice for private material. Practical uses abound: proofreading your own writing (the ear catches what the eye skips), accessibility testing of content, language-learning pronunciation practice with native voices, and hands-free reading of long articles. Because there are no accounts, no quotas, and no per-character billing, it also works as a dependable fallback when paid TTS APIs are down or over budget.

Which related tools should I use next?