Voice, audio and music with AI

In short

Transcription is the most useful and underrated use: meetings, interviews and lectures become searchable, summarisable text.
Text-to-speech is now almost indistinguishable from a human voice: ElevenLabs is the benchmark.
Voice agents answer the phone, book, route calls. Useful, but they must be designed carefully.
Cloning a voice requires the person’s consent, and generated audio that passes as a real person’s authentic voice (a deepfake) must be disclosed.

From voice to text: transcription

Automatic transcription is now high quality, with speaker recognition and punctuation. You’ll find it:

in video-call apps (Meet, Teams, Zoom) and phone recorders;
in assistants: upload an audio file to ChatGPT, Claude or Gemini and ask for a transcript and summary;
in dedicated tools for long work (podcasts, interviews, hearings).

The value isn’t the text itself, but what you do with it next.

Tell people before recording

Recording and transcribing a meeting with outside people requires informing them (and, where required, their consent). Transcripts contain personal data: treat them as such.

From text to voice: speech synthesis

With ElevenLabs (and similar tools) you paste a text and pick a voice: in seconds you get natural audio, with believable intonation and pauses. Typical uses:

audio versions of articles and newsletters (accessibility, listening on the go);
voice-over for videos, online courses, tutorials;
dubbing: the same video in several languages, keeping a voice similar to the original;
messages for switchboards and hold lines.

Practical tips: write for the ear (short sentences, no long lists), specify pronunciation of acronyms and foreign names, always listen to the whole file before publishing.

Cloning a voice: possible, but not casually

A few minutes of recording are enough to create a very faithful copy of a voice. It’s useful (a professional who “records” videos without reading them all) and dangerous (scams with cloned voices of relatives and executives are rising).

Minimum rules:

only your own voice, or that of someone who gave explicit, documented consent;
serious platforms ask for verification: it’s a safeguard, not a nuisance;
if you publish generated audio that sounds like a real person’s authentic voice, disclose it (AI Act, Art. 50: it’s a deepfake); otherwise disclosure is good practice;
in Italy, Law 132/2025 punishes spreading AI-falsified content that causes harm.

Voice agents: AI on the phone

A voice agent combines transcription, a language model and speech synthesis in real time: it understands what you say, decides what to reply, speaks. It can answer FAQs, book appointments, route calls, collect request details. Platforms like ElevenLabs offer no-code configurable agents; other companies, like Sesame, are experimenting with voice agents that use tools and computers to complete tasks.

To do it well:

the agent must say upfront it’s an AI (transparency duty);
it must be able to hand over to a person at any time;
limit it to clear tasks, with a controlled knowledge base (see the RAG guide);
record conversations and spot-check them, respecting privacy.

Music: Suno and songs on demand

With Suno you describe a track (“upbeat jingle for a bakery, acoustic guitar, 30 seconds, lyrics about the morning smell”) and get a complete song, with vocals and instruments. With Suno Studio you can edit sections, swap instruments, extend or shorten.

It works well for jingles, video backgrounds, musical prototypes, games. But beware:

commercial-use rights depend on your plan and can change: read them before using a track for a client;
professional musicians use these tools mostly to explore ideas, not as the final product;
don’t ask for tracks “in the exact style of” a living artist by name: besides legal issues, the result is almost always worse than describing the style.

In practice

Start with meeting transcription: the most immediate time saver.
For voice-over, try speech synthesis on a short video and compare with your own voice.
For voice agents and clones, start from the rules (transparency, consent, human handover), then the technology.
For music, read your plan’s terms before any commercial use.

Sources

Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.