How do modern text-to-speech systems let you control a voice's tone and pacing?

## Where the control moved Text-to-speech steering used to live in markup. The script was wrapped in XML-style tags that set speaking rate, pitch, volume, emphasis and explicit pause lengths, plus phoneme tags to force a pronunciation. That design fit engines that assembled speech from units or parameters, where each knob mapped to something mechanical inside the synthesizer. Neural voice models do not work that way. They generate prosody holistically from the text and from whatever conditioning they are given, so a global description of *how this should sound* is a better fit than dozens of local tags. As of mid-2026 the dominant interface is a free-text style instruction alongside the script — "warm, unhurried, slight smile", "crisp and factual, no filler", "apologetic but not grovelling" — with providers exposing it as a separate instructions field or as part of the voice configuration. Several leading models have explicitly dropped support for the old break and prosody tags. ## The three controls that matter today **1. Style instruction (global).** Sets register, emotion, pace and persona for the whole utterance. This is prompt-like: concrete sensory direction outperforms adjectives. "Like a pharmacist explaining a dosage to an anxious patient — slow, clear, reassuring" produces a more consistent result than "friendly". **2. Inline tags (local).** Markers placed in the script to trigger a specific behaviour at a point — a laugh, a sigh, a whisper, a beat of hesitation. The available set is model-specific, and unknown tags are a real hazard: depending on the model they are ignored or spoken aloud. **3. Pronunciation control.** For drug names, product names, surnames and acronyms, the reliable levers are a phonetic transcription (IPA is the common notation) or a respelling in ordinary letters that happens to synthesize correctly. A pronunciation dictionary applied across your whole script keeps this consistent instead of fixing terms one at a time. ## Why this matters in a voice product Delivery *is* content in speech. The same sentence read briskly and read slowly communicates different things about urgency and care. In a drive-thru or a support line, an over-cheerful read of an apology is worse than a flat one. Because the style instruction is free text, it can also be varied per situation — a confirmation read briskly, a price correction read more carefully — without rewriting the script. ## Practical gotchas - **Silent no-ops.** If the model ignores a control you assumed worked, nothing errors. The output just sounds ordinary. Audition every change. - **Tags read aloud.** The worst failure is a model speaking a literal tag to a customer. Validate scripts against the model's supported set and strip unknown markup. - **Numbers, dates and units.** "15 mg" and "1/2/26" are ambiguous to a synthesizer. Normalize them in text before synthesis rather than hoping the model guesses your locale. - **Instruction drift over long text.** A style instruction is strongest early; on a long passage the delivery can regress toward neutral. Synthesize in shorter chunks, restating the style, and stitch the audio. - **Non-determinism.** Neural TTS may not produce identical audio for identical input. If you need a fixed asset — a legal disclaimer, a brand line — synthesize it once and cache the audio file rather than regenerating it per request. - **Latency in a live loop.** In a conversational agent, synthesizing sentence by sentence lets the first words play while the rest is still being generated. Style instructions apply per request, so keep them consistent across chunks or the voice will shift mid-reply. ## Interview framing The point to make is not that markup is dead in every product — plenty of deployed systems still use it, and some providers still accept it. It is that the *locus of control* moved from declarative per-token markup to a natural-language description of delivery, because that is what neural voice models condition on well. Say which generation of tooling you are describing, and note that the details are provider-specific and change quickly, so the durable skill is auditioning output and building a pronunciation dictionary — not memorizing one vendor's tag list.