Text to speech Turning text into a voice, and the consent and licence questions that come with it.
Short answer. Text to speech (TTS), or speech synthesis, turns written text into spoken audio. It is the reverse of speech recognition: the task is to produce a waveform that matches the text in a chosen voice. Jurafsky and Martin describe modern systems as trained on tens of thousands of hours of speech from thousands of talkers, and some can imitate a new voice from about three seconds of sample audio, which is why consent matters. Quality is judged by listeners on a mean opinion score from 1 to 5. Open engines such as Piper run locally and list a Hindi voice, but each voice carries its own licence.
01
How it works
Older systems were built from one talker’s recordings, so a second voice meant a second recording project. Modern systems train one model on a very large amount of speech from many talkers, then steer it toward a voice using a short sample. Jurafsky and Martin describe a current approach in which a language model generates discrete audio tokens, produced by a neural codec, from the text and a few seconds of the target voice; the tokens are then turned back into a waveform.
Smaller local engines work in stages. Piper, for example, embeds espeak-ng to turn text into phonemes, and its voices are trained with a model called VITS and exported to run in ONNX Runtime.
02
Local or cloud
Option
Where the audio is made
Notes
Local engine such as Piper
On your own machine
Voices are files. Piper lists voices in more than 35 languages, including Hindi (hi_IN), Malayalam, Nepali and Telugu
Cloud voice service
In the provider’s cloud, so the text is sent out
Check retention, logging and whether the voice may be used commercially
03
Judging a voice for an office
Listen: quality is rated by people. Listeners score samples from 1 to 5 to give a mean opinion score, and two systems can be compared on the same sentences.
Check intelligibility automatically by running the audio through speech recognition and computing the word error rate, as the textbook suggests.
Test the hard cases: Indian names, rupee amounts, dates and mixed Hindi-English sentences.
Measure the delay before the first audio, which matters for any live conversation.
Listen to long passages, not only a single sentence; fatigue shows over minutes.
04
Consent, disclosure and licences
Do not imitate a staff member’s, client’s or public figure’s voice without their written consent. Systems that can imitate a voice from three seconds of audio make this a policy matter, not a technical one.
Label synthetic audio, in line with your AI disclosure policy.
Read each voice’s licence. Piper’s voices page says Piper is intended for personal use and text-to-speech research only, and that some voices have restrictive licences that must be reviewed.
Keep a log of what was generated and by whom.
05
Where ANVE stands
ANVE’s podcasts maker writes the script on the box, but narration uses a cloud voice today, which means the text is sent to that service. Local narration is not on ANVE yet. A voice front desk that answers callers in Hindi or English is one of the agents ANVE builds with a paid pilot, on AnveVoice, whose replies start a median 487 milliseconds after a caller stops speaking, measured end to end; it is not a shipped product.
06
Questions people ask
Is text to speech the same as voice cloning?
No. Standard TTS uses a stock voice. Newer systems can imitate a voice from about three seconds of sample audio, according to Jurafsky and Martin, and that needs the speaker’s consent.
Can text to speech speak Hindi offline?
Piper, an open engine that runs locally, lists a Hindi voice (hi_IN). Quality varies by voice, so listen to samples and check the licence before relying on one.
How is TTS quality rated?
By listeners, who score samples on a 1 to 5 mean opinion score. Speech recognition can also be run on the output to check that it is intelligible.
What is the difference between TTS and a voice agent?
TTS only speaks the text it is given. A voice agent also listens, using speech to text, and decides what to say, so its response time depends on all of those steps.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
