Speech Synthesis: Choosing a Voice Is Choosing a Commitment

Text to speech is the easiest capability in this whole syllabus to underestimate. It sounds like a solved utility — feed in a string, receive audio — and the technical part genuinely is straightforward. What makes it worth a note of its own is that synthesis is the one AI workload where the output is a persona, and personas carry consequences that transcription and extraction never do. This note covers how synthesis is controlled, the two voice paths available, the avatar extension, and the responsibilities attached.

From string to speech, and the controls in between

Synthesis converts input text into humanlike speech produced by neural voices. Left alone it makes reasonable default choices about pace, emphasis, and pronunciation. Those defaults are fine for a sentence and inadequate for anything longer, because natural speech is not uniform — a phone number is read differently from a poem, and a product name is not pronounced the way its spelling suggests.

The control surface for this is Speech Synthesis Markup Language, usually called SSML. It is a markup layer wrapped around the text that lets you adjust pitch, pronunciation, speaking rate, volume, and more. Conceptually SSML is to synthesis what layout is to a document: the plain content is only half the message, and the presentation carries the rest. A learner who has never seen SSML tends to assume synthesis quality is fixed by the voice, when in practice a large share of perceived quality comes from directing the voice properly.

Two voice paths, two very different commitments

The first path is the standard voices, a large gallery of ready-made options you can audition and adopt immediately. There is no lead time, no data collection, and no approval process. For the great majority of scenarios — reading articles aloud, navigation prompts, notifications, accessibility features — this is the correct answer, and reaching past it is over-engineering.

The second path is a custom voice, built so that it is recognisable and unique to a brand or product. Like custom speech models, custom voices are private and can constitute a competitive advantage. This is the path a media company takes to keep a consistent narrator across thousands of assets, or a brand takes when the voice itself is part of the identity.

The distinction that matters for identifying workloads is not audio quality. Standard voices are already excellent. It is ownership of an identity. A custom voice implies that a real person's vocal likeness is being reproduced, and that changes the nature of the project from a technical integration into something involving consent, contracts, and governance.

Avatars: the same commitment, made visible

The avatar capability extends synthesis into video, converting text into a digital video of a photorealistic human speaking with a natural-sounding voice. It can be produced ahead of time or generated in real time, it can be driven through an API or built without code, and its language coverage follows the same coverage as text to speech. You can pair it with a range of standard voices.

Everything true of a synthetic voice becomes more acute with a synthetic face. Microsoft's own framing keeps the responsible-AI condition attached to the capability rather than in a footnote: the feature exists to deliver lifelike synthetic talking avatar videos while adhering to responsible AI practices. That coupling is deliberate and it is examinable in spirit — the guardrails are part of the description of what the product is.

The obligations that travel with a synthetic person

Three responsibilities recur and are worth memorising as a set. Voice talent must give informed consent — the person whose voice is being reproduced must know and agree. Listeners should be able to tell that a voice is synthetic, which is the disclosure principle; deceiving people about whether they are hearing a human is the core misuse. And access to the most sensitive of these capabilities is deliberately gated rather than open to anyone with a subscription.

None of this is legal decoration bolted on afterwards. It exists because a convincing synthetic voice is directly usable for impersonation and fraud, and because the harm lands on a third party — the person whose voice was cloned — who is not a party to your project at all.

Mistakes to avoid

• Reaching for a custom voice by reflex. Ask what the standard gallery cannot do; usually the honest answer is nothing.
• Judging synthesis by the raw voice alone. Without SSML direction, good voices still read badly.
• Treating consent and disclosure as paperwork. They are design requirements that shape the product.
• Assuming an avatar is only a rendering step. Adding a face raises the stakes rather than merely adding polish.

What to carry forward

Synthesis has a technical half and an ethical half, and the second is the one that decides project scope. Standard voice or custom voice is really a question about whether you are borrowing an identity or creating one — and creating one brings consent, disclosure, and access review along with it.