Even if the voice is human-sounding to near perfection, even if it’s cleverly programmed to say “um,” and to follow a particular tone that one expects from conversational speech. Even if it’s injected with just the right amount of valley girl intonation or cool guy grit, you can hear it in the spaces between the words. You can tell it has no attachment to what it expresses, no subtext, and no voice of its own. The computer is speaking, but those words are not its words; it’s not thinking them up, and that void is where its soulful conversational spark goes to die.
However, this perception is changing with the current state of text to speech (TTS) technology. With newer TTS models, instead of just being read to, you can hear information conveyed in a more authentically human rhythm, prosody, and even voice tone. As a result, you can chat with a machine more conversationally, since it almost sounds like the machine is talking with you, and less transactionally, pulling responses from a response box.
Making AI Sound More Natural
Modern TTS systems generally use neural networks to analyze written text and predict elements such as pronunciation, pacing, emphasis, and intonation before generating the corresponding audio waveform. Newer models can also use contextual information from surrounding words and sentences to produce more natural transitions and expressive speech. However, these systems still have limitations. They may mispronounce unfamiliar names or technical terms, interpret ambiguous wording incorrectly, or produce inconsistent emotion and emphasis, particularly across longer passages or less-supported languages.
These developments can affect how generated speech is perceived. If the tone of a voice shifts appropriately with the content, it can sound more natural and may be easier to follow. It can help establish the credibility of a machine-generated persona, especially when users are expected to spend a lot of time listening to it speak.
More Ways to Use TTS
Text-to-speech technology has a number of useful applications. An application might provide it as a feature to read an article to a user while they multitask, or to generate realistic-sounding voices for videos and screen recordings, or to provide a way for users to transform written content into audio versions, or even to make an AI voice for their applications.
ElevenLabs provides an example of the direction in which text-to-speech technology has been evolving. It works in over 70 languages and aims to create an AI voice that uses the right rhythm, tonality, and intonation for the given text.
For developers, more natural-sounding voices can give them another way to shape the user experience in AI products. The vocal quality of a model can add a new experience and allow applications to present information in a way that can feel more conversational than text alone.
Language coverage is another important aspect of this model creation. Broad language coverage can also help creators and developers build voice projects for audiences across different regions and languages.
A Different Way to Interact With AI
A lot of the attention in artificial intelligence circles has been around what can be built: text and code generators, image creators, and more. Adding generated speech introduces another mode of interaction.
When a machine-learning system talks with expression and inflection, it gives users another way to engage with the output. They don’t just need to have it read back to them. The expansion of text-to-speech covers more ground than painting letters and words on a computer screen and having them parroted back to a user. It is supporting more conversational ways for people to interact with artificial intelligence.
