Synthetic speech crossed the naturalness threshold a while ago. A listener hearing a modern system read a paragraph generally cannot tell it is synthetic, and the remaining vendor differences on raw voice quality are small enough to be a matter of taste.
What still differs, and differs a lot, is control and latency. Those determine whether a voice application works.
What changed in 2026
- Latency became the competitive axis. As voice agents proliferated, time to first audio chunk mattered more than incremental naturalness gains.
- Emotional and prosodic control improved. Directing a voice to sound apologetic, excited, or measured became more reliable rather than a coin flip.
- Consent frameworks tightened. Requirements around cloning a real person's voice — documented permission, and disclosure — firmed up in both platform policy and law in several jurisdictions.
- On-device synthesis got usable. Local voices reached a quality level suitable for accessibility and offline applications.
What to evaluate
| Dimension |
Matters most for |
How to test |
| Time to first audio |
Conversational agents |
Measure at your prompt lengths |
| Naturalness |
Everything; largely solved |
Blind listening test |
| Pronunciation control |
Names, jargon, addresses |
Test with your actual vocabulary |
| Emphasis and pacing control |
Narration, explanations |
Try directing it and see if it obeys |
| Emotional range |
Character work, empathetic agents |
Task-specific listening |
| Consistency over length |
Audiobooks, long documents |
Listen to a full chapter, not a sample |
| Language and accent coverage |
International products |
Test each locale you serve |
| Cost per character |
High volume |
Compare at your actual volume |
Pronunciation is the item that generates the most user complaints and gets the least attention in evaluation. Every real application has names, product terms, or acronyms the model will mispronounce, and how easily you can correct that — through markup, a lexicon, or phonetic spelling — is a practical differentiator that no demo reveals.
Consistency over length is the audiobook trap. A system that sounds wonderful for thirty seconds can develop distracting rhythm patterns over an hour. Sample-length evaluation systematically misses this.
Latency in conversation
For a voice agent, the number that determines whether the interaction feels natural is time to first audio. Humans expect a response to begin within a few hundred milliseconds; beyond roughly a second the pause reads as a malfunction and people start talking over it.
That budget covers everything: transcribing the user, generating a reply, and beginning synthesis. Streaming synthesis — starting audio before the full text is generated — is what makes it achievable, and it requires the whole pipeline to stream rather than just the final stage. The integrated approach is covered in realtime voice API.
Consequently the highest-quality voice is frequently the wrong choice for conversation. A slightly less expressive voice that starts in two hundred milliseconds beats a beautiful one that starts in nine hundred.
The consent question
Cloning a voice technically requires seconds of audio. Doing it legitimately requires permission from the person whose voice it is, documented, with scope. That applies to employees, to voice actors, and to public figures regardless of how much of their audio is publicly available.
Platforms enforce this to varying degrees and law is tightening in several jurisdictions. The practical rule is straightforward: if you cannot produce written permission from the voice owner, do not clone the voice. The downstream misuse risk is covered in voice cloning scam protection.
Common mistakes
- Evaluating on thirty-second samples. Misses consistency problems entirely.
- Choosing maximum quality for conversation. Latency dominates perceived quality there.
- No pronunciation plan. Names and jargon will be wrong and users will notice.
- Cloning without documented consent. A legal exposure, not a grey area.
- Testing one language when you ship several. Quality varies substantially by locale.
FAQ
Can listeners tell it is synthetic?
Usually not from a short sample of a good system. Over longer passages, subtle rhythm patterns can give it away.
How do I fix a mispronounced name?
Most systems support pronunciation markup or a custom lexicon. Check that support before committing; it varies considerably.
Is on-device synthesis good enough?
For accessibility, navigation, and offline use, yes. For expressive narration, cloud systems still lead.
Do I have to disclose synthetic speech?
Increasingly yes in some jurisdictions and contexts, and it is good practice regardless. Disclosure costs little and builds trust.
Where to go next
For the input side, read speech-to-text compared. For building conversational systems, realtime voice API, and for the fraud angle, voice cloning scam protection.