Text-to-speech

Contemporary TTS is a two-part neural pipeline. An acoustic model turns text into an intermediate representation, typically a mel-spectrogram, and a vocoder turns that representation into a waveform. Some newer systems collapse both steps into a single end-to-end model. Either way the practical distinction for a voice agent is whether the model can stream: emitting the first audio chunk after a few input tokens rather than after the whole sentence.

Naturalness and controllability trade against each other. Highly expressive models produce convincing prosody but are harder to steer, and they occasionally mispronounce domain-specific words — product names, drug names, street names, surnames. Most production deployments therefore pair a good model with a pronunciation dictionary that forces the correct rendering of the words that matter to that business.

Sample rate matters more than it appears to. A 24 kHz or 48 kHz synthesis destined for a phone call gets downsampled to 8 kHz narrowband anyway, so audio tuned for headphone listening can sound worse on the PSTN than audio tuned for telephony. Testing a voice over a real phone call, not a browser preview, is the only reliable check.