When a navigation app reads out “turn left in 200 meters,” or an e-reader narrates a book aloud, that is text-to-speech. The system takes written words, works out how they should be pronounced and phrased, and generates the audio. Older systems stitched together recorded snippets and sounded robotic. Neural systems such as DeepMind’s WaveNet generate the sound wave directly and are much closer to human speech.
TTS is the mirror image of speech recognition, which turns audio into text. A classic voice assistant chains the two: recognize what you said, have a language model write a reply, then speak it. Newer multimodal models can take and produce audio within one model, which can cut delay and keep tone of voice. Businesses use TTS for phone support lines, accessibility features, audio versions of articles and dubbing.
Why it matters: synthetic speech is now good enough that listeners may not realize it is synthetic. That helps accessibility and localization, but it also enables voice cloning without permission, which is a common ingredient in phone scams and impersonation. Responsible use means getting consent from anyone whose voice is copied and telling listeners when they hear a synthetic voice.
In episode 43 at 13:19, Liquid AI’s Maxime Labonne describes an audio model that handles “speech to text, but also text to speech or speech to speech,” and suggests it could power an in-car assistant. In episode 47 at 31:59, Google DeepMind’s Paige Bailey says the audio in a demo app came from Gemini text to speech, saved ahead of time so it loaded instantly; the episode was sponsored by Galileo (sponsor read). In episode 20 at 35:25, Smartling’s Olga Beregovaya points to synthetic voices and “by consent, voice cloning” as opportunities that need care about ethics. Episode 20 aired in season two, which Galileo sponsored.
Hear it from the guest
“It can do speech to text, but also text to speech or speech to speech or text to text.”
“The audio snippets were all using Gemini text to speech”
“Potentially of course by consent, voice cloning”
Quotes lightly edited to remove filler words.
Sources
- Google Cloud: Cloud Text-to-Speech basics — Official product documentation describing an API that converts text or SSML input into natural-sounding synthetic speech as audio files such as MP3; checked October 24, 2026.
- van den Oord et al.: WaveNet: A Generative Model for Raw Audio — Introduces a neural network that generates raw audio one sample at a time and, applied to text-to-speech, was rated more natural than earlier systems; checked October 24, 2026.
