Text-to-speech: understanding synthetic voices and using them well on the phone

Speech synthesis covers all the techniques that produce speech from text. The process breaks down into steps that the W3C’s SSML specification, a standard markup language for speech synthesis, describes clearly: analysing the structure of the text, normalising it (spelling out “12:30” or “Dr” in full words), converting words into phonemes, calculating prosody, that is intonation and rhythm, and then producing the sound signal.

Each step can trip up. Normalisation has to know whether “1/2” is read as “one half” or “one in two”; phoneme conversion has to know how to pronounce a street name or a brand; prosody has to tell a question from a statement. That is the very purpose of SSML: to let you tell the synthesis engine about a pronunciation, a pause, a volume or a speaking rate.

Early approaches stitched together fragments of recorded voice, which produced audible transitions between the pieces. Parametric methods then modelled the voice using acoustic parameters, giving a more uniform but more mechanical sound. Today’s neural voices learn to produce the signal from large quantities of recordings and deliver much more natural intonation.

This progress has a downside: a very natural voice can make the listener forget they are talking to a machine. That is precisely what European regulation seeks to prevent, and it is also a matter of trust for your business: a client who finds out afterwards that they were talking to an AI may feel deceived.

On the phone, the ear only gets compressed and sometimes degraded sound. A voice that sounds pleasant in a demo on a good speaker can become tiresome on a mobile line. The criteria that matter are therefore intelligibility, a moderate pace, pauses in the right places and clear articulation of numbers, because a misheard phone number or appointment time defeats the purpose of the call.

Test the voice with your own content: your company name, your staff’s names, your addresses, your prices, the technical terms of your trade. Listen for latency too, that is the delay between the end of the caller’s sentence and the start of the reply: a perfect voice that answers after a long silence gives the impression that the line has dropped.

Article 50 of the European Union’s Artificial Intelligence Act has applied since 2 August 2026. According to the European Commission, providers of AI systems intended to interact directly with people, such as conversational agents, must ensure that those people are informed, at the latest at the time of the first interaction, that they are interacting with an AI, unless this is obvious to a reasonably well-informed person.

The same article requires providers to mark AI-generated audio content in a machine-detectable way. According to the Commission, a grace period until 2 December 2026 applies, for this marking obligation only, to systems placed on the market before 2 August 2026. Deepfakes, which for example imitate the voice of a real person, must be disclosed as such. For phone use, the practical rule is simple: announce within the first few seconds that the caller is speaking to an AI.

In its white paper on voice assistants, the CNIL, France’s data protection authority, points out that the voice is a biometric characteristic that can make it possible to identify a person. Cloning an employee’s voice to make it the voice of your switchboard is therefore not a trivial step: it is wise to have the applicable legal framework checked, obtain their consent and plan what will happen if they leave the company.

A generic synthetic voice, clearly presented as such, avoids these difficulties. It also has the advantage of creating no ambiguity for callers, who will not think they are speaking to a specific person on your team.

Alyas uses speech synthesis to answer the phone 24 hours a day using the information you provide, and announces at the very start of each call that it is an AI and that the conversation is recorded. Every call produces a transcript and a summary that you can consult in your client area, which helps you spot the moments where the caller misunderstood or asked for something to be repeated. The ten-day free trial lets you hear it on your own calls.