Voices CTO argues better AI speech depends on better training data

Voices CTO Dheeraj Jalali argues that the next gains in AI speech will come from better-targeted training data, not simply more audio. His case is useful, but it comes from a company that sells the licensed recordings it recommends.

Watch Desk analysis

What happened

In an interview with The Tech Buzz, Jalali distinguishes speech recognition from text-to-speech. He says noisy, real-world recordings can be useful for training recognition systems, while expressive speech generation needs recordings that capture things such as cadence, emotion and naturalness.

Jalali describes Voices’ process: session directors coach contributors, recordings are checked against acoustic measures including signal-to-noise ratio and clipping, and files are delivered with transcripts, timestamps, speaker IDs and metadata. He says contributors agree to the intended use, including AI training. The interview does not establish how these methods compare with alternatives, and Jalali did not answer whether contributors can revoke consent after training.

Why it matters

The distinction helps explain why a model that can clone a voice from a short sample does not mean its underlying training data was equally quick or cheap to assemble. For teams building speech products, the relevant question may be whether their data covers the voices, languages and situations the product needs, not just how many hours sit in the collection.

It also puts consent and documentation in the practical frame: who supplied the recordings, what use they agreed to, and what information travels with the files. Those are important questions for buyers, even when the answers come from a vendor making its own case.

Our read

Jalali makes a persuasive argument that specialised speech data can matter more than raw volume for particular tasks. But the interview offers a vendor’s account, not a comparison showing that licensed studio recordings consistently outperform other approaches. The unanswered question about withdrawing consent is not a footnote; it is part of what a durable consent model needs to explain.

For developers, ask suppliers how recordings are sourced, what uses are covered, what metadata is supplied and what happens when a contributor changes their mind. A tidy waveform is nice. A clear agreement is rather more useful when someone asks where it came from.

What to watch

• Whether model builders disclose specific training-data sources and licensing arrangements.
• Whether speech-data providers explain how consent can be withdrawn after training.
• Whether independent comparisons show where directed, licensed recordings improve speech quality over other data.

Discussion spark: Should consent to train a voice model include a right to withdraw after training, even if honouring it could mean rebuilding the model?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.