Text to Speech

Text-to-speech converts written text into synthesized spoken audio. See Speech to Text for the inverse direction (audio to text); the two are often paired in voice-interface applications but solve opposite problems with largely different underlying techniques.

How modern TTS works

Most current TTS systems are neural: a model predicts an intermediate acoustic representation (a mel-spectrogram) from text, then a vocoder (another neural model) converts that into a raw audio waveform. This has mostly displaced older concatenative synthesis (stitching together pre-recorded speech fragments) and formant synthesis (rule-based acoustic modeling), both of which tend to sound noticeably more robotic and less natural than modern neural voices.

SSML: controlling pronunciation and prosody

<speak>
  Welcome to <say-as interpret-as="characters">SSML</say-as>.
  <break time="500ms"/>
  The meeting is at <say-as interpret-as="time">3:00pm</say-as>.
  <prosody rate="slow" pitch="+2st">This part is spoken slowly.</prosody>
</speak>

Speech Synthesis Markup Language (SSML) is a widely supported standard for controlling how text is spoken — pronunciation of ambiguous terms, pauses, emphasis, rate, and pitch — rather than relying purely on the engine's default interpretation of plain text.

Choosing a TTS provider/approach

Options range from cloud APIs (Amazon Polly, Google Cloud Text-to- Speech, Microsoft Azure Speech, ElevenLabs) offering a wide range of high-quality voices with minimal setup, to self-hosted open-source models (Coqui TTS, Piper) for offline use, data privacy, or cost control at scale. The trade-offs are typically: cloud APIs give the best voice quality and least engineering effort but recurring cost and a network dependency; self-hosted models give control and no per-request cost but require more infrastructure and generally lower peak voice quality.

Common pitfalls

• Ignoring pronunciation of domain-specific terms — acronyms, names, and jargon are frequently mispronounced by default; use SSML or a custom pronunciation lexicon where the engine supports one.
• No handling for unsupported languages/locales — a TTS engine's voice quality and language coverage varies significantly; verify actual support before assuming a target locale is well-covered.
• Treating synthesized speech as accessible by default — TTS quality alone doesn't guarantee a good experience for someone relying on it as their primary interface; test with real usage patterns and see Screen Reader for the specific assistive-technology context TTS often serves.
• Not caching or reusing generated audio — resynthesizing the same static text repeatedly wastes cost and latency versus caching the audio output for reuse.

Learn more

• Speech to Text for the inverse conversion.
• W3C SSML specification
• Screen Reader for a common consumer of synthesized speech.