Text to Speech Online — Neural Voices

132 voices across 34 languages in total, from five sources: Microsoft cloud voices, our own server voices, HD voices rendered by your browser, and studio voices from OpenAI and Yandex billed in tokens. For the language of this page there are 56 voices; every language has its own set — the “Languages and available voices” table below shows how many.

Your text is sent to our server, which connects to Microsoft's neural speech service via a secure WebSocket to generate audio. The text is used only for the duration of the synthesis request.

The limit depends on the account, so there is no single number: without signing in, 3 000 characters per synthesis and 3 000 per day; with Premium, 30 000 per synthesis and 100 000 per day. Your current allowance is always shown under the input, and the field itself truncates text longer than your limit. For longer texts, split them into parts that fit your limit, generate each separately, then join the MP3 files with our audio joiner.

Pauses, yes — and precisely: write [pause 3s] between phrases and the service inserts a pause of exactly that length (0.1 to 30 seconds — handy for workouts, meditation, step-by-step instructions); /pause/ gives a short pause. Emphasizing a single word with markup is not supported on the standard voices — for emotional delivery pick a speaking style in the settings or a premium voice with 'Vibe'.

MP3 for anything you are going to publish or send, WAV if the speech is going into an editor and will be cut, layered or processed further. The speech itself arrives as MP3 at 24 kHz, which is the sample rate the voice was generated at; choosing WAV on download converts that in the page rather than fetching a better original, so it does not add detail that was never there — what it does is stop a second round of lossy encoding when you re-export from your editor. For a podcast or a video, MP3 at the higher bitrate is indistinguishable and a tenth of the size.

Auto-detection works best with sentences (3+ words). For very short text or mixed-language content, select the language manually from the dropdown. The tool detects language by analyzing Unicode script ranges and common word patterns.