Voice and speech tools

What the category contains

Text to speech. Turning writing into audio. Free and built into every operating system; better voices available from paid services and from open models.

Voice cloning. Producing speech in a particular person’s voice from a sample, sometimes only seconds long.

Speech to speech. Changing an existing recording to a different voice, keeping the timing and delivery.

Dubbing. Translating a recording and speaking it in the new language, sometimes matching the original speaker’s voice.

The technology has moved fast, and the honest summary is that the good hosted services are now hard to distinguish from a person on a short clip, and the free options are good enough for most practical uses.

The free path

Piper runs locally, produces natural speech in many languages, and is fast enough on a Raspberry Pi. This is the workhorse for accessibility, for reading documents aloud, and for anything you produce regularly.

Coqui TTS and similar projects offer more voices and cloning at the cost of more setup.

Your operating system. Built-in voices have improved considerably. For reading a document back to yourself, no installation is required at all.

Where these help

• Accessibility. For someone who reads with difficulty or not at all, having any document read aloud in a natural voice is transformative, and it is free.
• Proofreading. Listening to your own writing exposes errors the eye slides over. This is an old editor’s trick and it works better with a good voice.
• Learning a language. Hearing a sentence read correctly, on demand, at your own pace.
• Reaching people who do not read your written language but speak it: a large population in many countries, and one that written-only material excludes entirely.
• Producing audio versions of material at a cost that makes it worth doing for small audiences.

The consent problem, which is the important part

Voice cloning creates a genuine harm that has already happened many times. Recorded voices are used in fraud: the call that sounds like a family member in trouble, the instruction that sounds like a manager approving a payment. A few seconds of audio from a public video is enough for some tools.

Three things follow.

Do not clone a voice without the person’s explicit permission. Reputable services require an attestation, some require a spoken consent recording, and it remains the user’s responsibility. Several jurisdictions now have laws about voice likeness, and the rules differ.

Assume other people’s voices can be cloned. The practical defence for your family and your workplace is a shared phrase or a callback rule for anything involving money or urgency. That advice sounds excessive until it has happened to someone you know.

Disclose synthetic voice where the listener would care. In advertising, in journalism, in anything a person might rely on. Some platforms and some regimes now require labelling, and the requirements are still forming.

Pronunciation, which is where most of the work goes

The gap between an acceptable read and a good one is almost entirely names, acronyms and foreign words. Every serious tool accepts a pronunciation override: a phonetic spelling, a dictionary of terms, or SSML markup that specifies how a word is said.

Building that dictionary once, for the twenty words your material keeps using, is an hour that improves everything you produce afterwards. It is also the reason a free local tool with a good dictionary often beats an expensive service used straight out of the box, and it is the part of the work no product page mentions.

Choosing a service

Beyond quality, three things separate them for real work:

• Languages and accents, particularly whether your language is supported at all rather than as an accented approximation.
• Control over delivery: pauses, emphasis, pace, pronunciation of names. On a long piece this is what turns an acceptable read into a good one.
• Commercial terms. Whether you may use the audio commercially, whether the voice is licensed for it, and what happens to the voice you cloned. Read this before recording a client’s audiobook, not after.

For anything you produce regularly, test the free local option first. The gap between Piper and a paid service is real and it is smaller than the gap between doing this at all and not doing it.