Advanced Studio: how each mode fits together

Single. It has the gentlest learning curve — type text, pick a voice, hit Generate. Most users only need Single to find their preferred voice and pacing. Move on to Dialogue when you need multiple characters, Timeline when you want to chain clips, and Audiobook when your project is too long for Single's 25,000-char cap.

Yes. Every Generate click in Single counts. Voice previews (the small play button next to a voice in the picker) do NOT count — those use a separate quota-free endpoint. Test as many voices as you want via Preview without burning chars.

Not directly. Dialogue muxes the conversation server-side and returns one merged file. If you need separate files per speaker, generate each speaker's lines in Single mode (switching voices between them), then add each one to Timeline for chaining. It's slower but gives you full control over the individual audio files.

Timeline assembles audio you already generated. Audiobook generates new audio from text. Use Timeline when you want to merge specific clips (e.g., chain three dialogue scenes plus a narrator intro). Use Audiobook when you want to give it a full chapter and get back a finished ZIP. They serve different jobs — Timeline is editor-style, Audiobook is batch-pipeline-style.

Audiobook splits long text into chunks (Azure's batch API caps each chunk at ~10,000 chars). The ZIP contains one MP3 per chunk plus SRT subtitles. If you want a single merged file, you can either run the chunks through Timeline (add them all + merge) or use an external tool like ffmpeg / Audacity.

Yes — Single and Dialogue both accept SSML. The most useful tags are <break time="1s"/>, <emphasis level="strong">, <prosody rate="slow">, and <mstts:express-as style="cheerful">. Audiobook auto-detects SSML too (as of May 2026). All tags are lowercase — <Break/> with a capital B won't work.

Each Audiobook submission writes a single row to your history with the title and total chars, not one row per internal chunk. The History table's Source column will label it 'Audiobook' so you can tell it apart from Single (web), Dialogue (studio), and API entries.

Azure's batch service typically completes in roughly half the audio's playback length. A 1-hour audiobook takes ~30 minutes. Watch the dashboard 'Recent batches' panel for live progress. The job runs server-side, so you can close the tab — you'll see the finished link on next visit.

HD voices (like en-US-Aria:HDLatestNeural) are higher-bitrate versions of standard neural voices — same expressive control via mstts:express-as, just better audio quality. DragonHD voices (like en-US-Andrew:DragonHDLatestNeural) are a newer-generation model from Microsoft that automatically reads emotion from the text itself. You write "She gasped in horror" and DragonHD delivers it dramatically without you needing the terrified style. The catch: DragonHD rejects mstts:express-as tags entirely, so manual style control is gone. Use HD when you want explicit control; use DragonHD when you want natural delivery from natural writing.

Three common reasons. First, your voice doesn't support that style — only certain voices have specific styles (the Studio chip picker shows what your selected voice supports). Second, you're on a DragonHD voice, which rejects express-as tags entirely. Third, you used the style outside an mstts:express-as wrapper. Audiobook auto-wraps the SSML envelope for you, so this only happens if you're writing raw SSML for the API.

Yes. The homepage /api/tts endpoint and the browser extension both build a full SSML envelope from your text. Drop in <break/>, <emphasis>, <prosody>, even mstts:express-as — they all work on the homepage if the underlying voice supports them. The 1,000-char free-tier per-generation cap still applies though.

styledegree controls how intense the emotional style is. Default is 1.0. Range is 0.01 (barely-there hint) to 2.0 (cranked up). For most narration the default works fine. Bump to 1.5 or 2.0 for dramatic moments (a battle cry, a grief explosion). Drop to 0.5 for subtle inflection (a hint of sadness under a brave face). Add it as an attribute: <mstts:express-as style="sad" styledegree="0.5">.

Azure caps `<break time>` at 10 seconds per tag. For longer silences, stack multiple breaks: <break time="10s"/><break time="10s"/><break time="5s"/> gives you 25 seconds. Or use mstts:silence at the top of your document to set a global between-sentence silence if you just want spacious pacing throughout.

Yes, on the Creator plan. You record 60 seconds of clean audio of yourself talking naturally (no background noise, normal pace), upload it from /dashboard?section=voice-clone, and the system trains a personal voice in about 5 minutes. Once cloned, your voice appears in the picker like any other voice. Creator includes 1 voice clone and 100,000 cloned characters per month. Quality scales with the recording — invest in 60 seconds of really clean audio and the clone will sound impressively close.

Different paths handle invalid SSML differently. Real-time TTS (Single, Dialogue) returns a clear error in the UI telling you what's wrong — usually with the line position. Audiobook batch is less forgiving — if one chunk has invalid SSML, that whole chunk fails and is skipped (the rest succeed). Test SSML in Single mode first if you're adding it to a long batch — much faster feedback loop.

Yes, three ways. Easiest: pick a Multilingual voice (look for "Multilingual" in the voice name) which can speak ~12 languages with one identity. Most precise: use <lang xml:lang="fr-FR">phrase</lang> inside any voice to switch languages mid-sentence for that span. Most flexible: use Dialogue mode and assign each language to a different voice — you get separate native speakers for each language.

Neural voices have a small amount of natural variation — the same text rendered twice won't be byte-identical even with the same voice and settings. Variation is normally subtle (different breath placement, slight rhythm differences). Larger differences usually come from context — the voice reads a question differently from a statement, an exclamation differently from a declaration. This is a feature, not a bug, and is part of what makes neural voices feel human.

A PRO/Creator feature that stores custom word → pronunciation mappings that auto-apply to ALL your generations. Add 'Camarath' → 'kəˈmɛrəθ' once, and every time you write Camarath in any Studio mode it gets pronounced correctly. Much cleaner than wrapping every instance in <phoneme>. Find it in the dashboard sidebar under Studio settings. Uses International Phonetic Alphabet (IPA) — there's a 'Hear it' button to audition before saving.

Azure's batch synthesis caps each chunk at ~10,000 chars to keep individual jobs fast. A 200K-char audiobook becomes ~20 chunks, each delivered as a separate MP3 in the ZIP along with SRT subtitles. To merge into a single MP3: extract the ZIP, then either (a) use Studio's Timeline mode to add each chunk and merge, or (b) use ffmpeg locally: `ffmpeg -f concat -i list.txt -c copy output.mp3` where list.txt contains "file 'chunk1.mp3'" lines. The ZIP keeps things flexible — chapter-level files are easier to edit and reorganize than one giant MP3.

Yes — that's exactly what Dialogue is for. Write your script as `Speaker: line of dialogue` (one per row). Assign each speaker a voice in the picker. Dialogue mode auto-generates the multi-voice SSML for you. Per-line style overrides are also supported via the small style dropdown next to each line. You only need to drop into raw SSML if you want effects beyond voice + style (custom pauses, prosody, etc.).

All three are generated from the same 24 kHz neural synthesis source. MP3 (default, 160 kbps) is universal and small. WAV is uncompressed PCM — same source quality, ~2.5x file size, ideal for re-editing in a DAW. OGG/Opus is the most modern, smallest at equivalent quality, but older browsers and devices don't support it. Pick MP3 unless you have a specific reason to use the others.