What the model is actually predicting
Reading a sentence aloud requires far more information than the sentence contains. Where does stress land, how long is each vowel, does the pitch rise at the end, where does the speaker breathe. None of that is written down, and a TTS model's real job is to supply it.
The standard architecture separates that job into two. An acoustic stage takes normalised text, often converted to phonemes first, and predicts a spectrogram: a time-frequency picture that encodes pitch contour, duration and timbre but is not yet sound. A vocoder stage then synthesises the waveform from that picture. The split matters in practice because the two stages fail differently. Acoustic failures sound like wrong emphasis, odd pacing, or a flat reading. Vocoder failures sound like buzz, metallic ringing or smeared consonants.
Speaker identity enters as a separate conditioning signal, which is why one model can hold many voices, and why a voice can be swapped without rewriting the text.
The controls that matter
Most systems expose some subset of the following, and they are not equally powerful.
- Voice selection. The largest single factor. Each voice carries its own pitch range, accent and default energy, and no parameter will make a calm narration voice shout convincingly.
- Speed. Usually a clean multiplier. Small changes read as naturalness; large ones read as tape speed.
- Emotion or style. Where offered, this is a strong lever, but it is a global one. It sets a mood for the whole request rather than a beat within it.
- Stability or variance. Higher stability gives repeatable, flatter reads; lower gives more expression and more risk of an odd take.
- Language or accent hints. Important for mixed-language scripts, where a model will otherwise apply one language's phonology to the other's words.
The control most people underuse is the script itself. Punctuation is a timing instruction. A comma is a short beat, a full stop is a longer one, and a paragraph break is longer still. Rewriting a sentence to be shorter changes delivery more reliably than nudging a speed slider.
Writing text that reads well aloud
Text normalisation sits in front of the whole pipeline and it is where most embarrassing errors originate. It has to decide that Dr. is doctor rather than drive, that 1420 is fourteen twenty rather than one thousand four hundred and twenty, and that read is present or past tense. It guesses from context and it is often wrong.
Practical rules. Spell out anything you care about: write fourteen twenty rather than 1420. Respell unfamiliar names the way they sound. Break long sentences at the points where a person would breathe. Avoid parentheses, which most systems either ignore or read as a stumble. And check acronyms, since a model may spell one out letter by letter or attempt to pronounce it as a word, with no way to predict which.
For anything longer than a paragraph, synthesise sentence by sentence and assemble with deliberate silences. This costs an extra step and buys three things: consistent prosody, the ability to re-roll one bad line without regenerating the whole take, and exact control over pause length, which is what actually makes narration sound edited rather than recited.
Where it sits next to the other audio tasks
Speech synthesis from text is one of four related jobs, and choosing correctly saves a lot of trial and error. Text to speech starts from a script and a chosen voice. Voice cloning starts from a script plus reference audio of a specific person. A voice changer starts from a recorded performance and only replaces the timbre, which is the right tool when the acting matters more than the wording. Lip sync starts from finished audio and a video, and moves the mouth to match.
The pipeline choice follows from what you already have. If the performance exists, keep it and convert. If only the words exist, synthesise them, and expect to spend most of your effort on punctuation and segmentation rather than on parameters.