Text-to-Speech
Definition
Text-to-Speech
Text-to-speech (TTS) turns written text into audible, human-sounding speech using neural voice models. It powers screen readers, IVR phone menus, audiobooks, and conversational agents on web, mobile, and phone, replacing studio audio with synthesised voices on demand.
Modern TTS engines combine linguistic analysis, prosody prediction, and neural vocoders trained on hundreds of hours of narrated audio.
The result sounds close to a human speaker, with adjustable pitch, pace, and emotion controlled through SSML tags or plain API parameters.
Outsourcing providers deploy TTS across contact-center automation, e-learning production, and accessibility tooling, trimming voice-over budgets while scaling multilingual coverage far beyond what a booth-based studio could deliver in the same week.
Key takeaways
- TTS synthesises speech from text using neural voice models, not pre-recorded clips.
- Cloud APIs from Google, Amazon, and Microsoft price TTS by characters processed.
- Contact centers pair TTS with speech-to-text and NLU for full voice-bot loops.
- SSML tags let developers control pitch, rate, pauses, and pronunciation per phrase.
- Accessibility law in many markets requires TTS-ready digital content for public bodies.
How it works
A TTS pipeline ingests raw text, normalises numbers and abbreviations, predicts phonemes and prosody, then a neural vocoder renders the waveform. Cloud vendors expose this behind a single REST call returning MP3 or WAV audio.
Under the hood, the system runs three stages. First, text normalisation expands “Dr.” to “Doctor” and “2024” to “twenty twenty-four”.
Second, an acoustic model, usually a Tacotron or FastSpeech variant, predicts mel-spectrograms. Third, a neural vocoder such as WaveNet or HiFi-GAN converts spectrograms into audio samples.
| Stage | Input | Output | Typical model |
|---|---|---|---|
| Normalisation | Raw text | Cleaned phonemes | Rule-based + LLM |
| Acoustic | Phonemes | Mel-spectrogram | FastSpeech, Tacotron |
| Vocoder | Spectrogram | Waveform | WaveNet, HiFi-GAN |
| Post-process | Waveform | MP3 / WAV | Codec encoder |
Google Cloud published its WaveNet TTS in 2018 and has since expanded to 380+ voices across 50+ languages, priced per one million characters processed.
Latency budgets shape architecture choices. Real-time voice bots need audio streaming within 300 milliseconds — batch workloads like audiobook generation trade latency for higher-fidelity vocoders and longer prosody planning windows.
Enterprise buyers weigh three variables: voice quality, language coverage, and per-character cost. A voice-bot with 40 language locales at $16 per million characters may lose to a leaner engine with 20 locales at $4, depending on traffic mix and market split.
Examples
Real deployments span consumer apps, accessibility products, and enterprise voice channels. The common thread is scale, TTS costs pennies per thousand characters — so it beats studio recording anywhere content changes weekly.
- Duolingo uses neural TTS to voice new lesson content across 40+ languages without booking voice actors for every update, a workflow shift the company documented publicly in 2023.
- Amazon Polly powers Kindle’s “read aloud” feature and thousands of IVR menus. AWS lists Polly among its flagship ML services with per-character pricing starting at $4 per million characters.
- Microsoft Azure Neural TTS ships inside Immersive Reader, used by tens of millions of students through Microsoft 365 Education for on-demand text narration.
- Descript and ElevenLabs offer voice-cloning TTS that podcast studios use to patch mispronounced words without re-recording the host.
Related terms
TTS sits inside a wider stack of speech and language technologies. The neighbours below share vocabulary, pipelines, or deployment patterns, worth reading alongside this entry.
- Artificial Intelligence (AI): the parent field that produces the neural models behind modern TTS engines.
- Machine Learning: the training approach used to teach voice models from paired text and audio.
- Large Language Model: often paired with TTS to generate and then speak conversational replies.
- Multimodal AI: systems that jointly handle text, audio, and images, the direction TTS is heading.
- Contact Center: the primary enterprise buyer of TTS for IVR and voice-bot deployments.
- Customer Experience: the outcome metric TTS quality most directly moves in voice channels.
- Automation: the broader category TTS falls under when replacing recorded voice-over work.
FAQ
How accurate does TTS sound compared to a real voice?
Top-tier neural TTS from Google, Amazon, or Microsoft passes casual listener tests in short clips. Longer narration still shows small prosody artefacts on complex sentences — which is why audiobook studios use TTS for drafts and humans for final reads.
Can TTS clone a specific person’s voice?
Yes, services like ElevenLabs and Azure Custom Neural Voice train a personalised model from 30 minutes of clean audio. Ethical deployment requires explicit consent from the speaker and clear disclosure to listeners.
How much does cloud TTS cost?
Public pricing sits between $4 and $16 per one million characters as of 2024, depending on voice tier and vendor. That works out to roughly $0.005 for a 300-character customer-service reply.
Does TTS support languages beyond English?
Google Cloud TTS covers 50+ languages, Amazon Polly around 40, and Microsoft Azure 140+ locales — so most major markets are served.
Outsource Accelerator connects buyers with vetted BPO partners deploying TTS inside contact-center stacks, start at Outsource Accelerator.
Independent
