Text-to-Speech
Turning text into speech that sounds human — where the remaining gap isn't the voice, it's knowing which word to stress.
When not to use it
- When a recording would do. For fixed content read many times, record a human once. It's better and cheaper.
- When precise emphasis carries the meaning. Legal readouts, safety instructions, anything where stressing the wrong word changes what was said.
- When latency is the product and you've chosen quality. A conversational agent that takes two seconds to start speaking has failed regardless of how good it sounds.
- On text with unmarked names, numbers, or jargon. It will mispronounce them confidently and you won't hear about it from users, they'll just leave.
Reach for something else instead
- Recorded audio — better, for anything fixed.
- Concatenative synthesis — old, constrained, and utterly predictable. Occasionally the right answer for safety-critical fixed phrases.
- Streaming TTS — when latency beats fidelity.
- Voice actors — for brand and long-form. The gap is prosody, and a person has intent.
Sources & further reading
- van den Oord et al. (2016), WaveNet: A Generative Model for Raw Audio — the paper that ended robotic speech, and was far too slow to ship.
- Shen et al. (2018), Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions — Tacotron 2; the two-stage shape most systems still use.
- Kong et al. (2020), HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis — how the quality got fast enough to be a product.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Evaluating on the demo sentence. Test with your actual text, including your names and numbers.
- Ignoring SSML, then wondering why the pauses are wrong. The markup is the control surface.
- Choosing the highest-quality voice for a conversational product and shipping the latency.
- Not overriding pronunciation for domain terms, which are precisely the words that matter.
- Assuming a natural voice implies natural delivery. Naturalness is solved; prosody isn't.
At a glance
FieldSpeech & Audio
Also calledTTS, speech synthesis
Pipelinetext → spectrogram → vocoder
Solvednaturalness
Not solvedprosody
Control surfaceSSML
DifficultyBeginner
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
TTS vs. a recording — for fixed content, record a human; it's better and cheaper. TTS wins when the text isn't known in advance.