Everything we have written
How to get good audio out of this, what it costs against the alternatives, where it will let you down, and what we broke on the way.
- 1,481labelled voices, licences printed
- 11catalogue languages
- 23Indian languages
- 35translation languages
- 30 minfree every month
- AudioSealwatermark on every render
Guides
How to do one job with the product, start to finish.
- How to turn text into speech that does not sound like a robot A practical text to speech guide: pick a voice, fix the pronunciation and pauses, render an MP3 or WAV, and know what you are allowed to do with the file.
- What an AI voice generator can and cannot do in 2026 An honest account of AI voice generation: what sounds human now, what still gives it away, what it costs, and where the law has caught up.
- Voice cloning, done legally and done well How voice cloning works, what a usable reference recording sounds like, what consent you need on file, and where cloning is the wrong tool.
- Speech to speech: keep your performance, change the voice How speech-to-speech voice conversion works, when it beats text to speech, and how to use an acted reference to get emotion out of a synthetic voice.
- Hindi and Indian-language text to speech that reads Devanagari properly Hindi, Tamil, Telugu, Bengali, Marathi and 18 more: how Indian-language text to speech works on Naadly, what it costs, and where it still struggles.
- Japanese text to speech, and the kanji problem How Japanese text to speech handles kanji readings and pitch accent, what Naadly uses, and how to check a render before it embarrasses you.
- A text-to-speech API you can read in one sitting Authenticate, list voices, render speech, stream it, and get told when a long job finishes - with Python and JavaScript examples and the real limits.
- Dubbing a video with AI: transcribe, translate, re-voice How to dub video into another language with AI - transcription, translation, voice matching, subtitles and the lip-sync problem nobody solves.
- Narrating an audiobook with AI without it sounding like one How to produce an audiobook with AI narration: voice choice, chapter splitting, pronunciation lists, consistency and what the retailers require.
- AI voice-over for YouTube and short-form video How to make AI voice-over for YouTube, Shorts, Reels and TikTok that fits the edit, keeps monetisation intact and does not sound like every other faceless channel.
- Voice-over for courses and training that survives version 7 How to narrate e-learning modules with AI so that updating slide 14 next quarter costs one line, not one studio day.
- AI voice for podcasts: intros, ads, corrections and full episodes Where synthetic voice belongs in a podcast workflow - and where listeners will notice - plus transcripts, ad reads and fixing a line without a re-record.
- Transcription and subtitles that are actually usable How to get an accurate transcript and clean SRT subtitles from audio or video, what accuracy to expect, and how to fix the parts machines always miss.
- Pronunciation, pauses and timing: the controls that matter How to make a synthetic voice say your product name properly, pause where you want, and hit a length - with punctuation, a pronunciation dictionary and per-line speed.
- Bulk and long-form: hundreds of lines without babysitting How to render hundreds of lines or a whole book with a CSV import, queued jobs and webhooks - and what to check before starting a job you cannot watch.
- Give Claude, Cursor and ChatGPT a voice with MCP How to connect Naadly to an MCP client so an assistant can list voices and render speech itself - setup, what the tools do, and the limits.
- What you may legally do with AI-generated audio Who owns AI-generated audio, when open voice models require attribution, what a client can ask you for, and how Naadly's licensing works on every plan.
- Open-source text to speech: what it costs to run yourself Piper, Kokoro, Whisper and friends: what open TTS models are good at, what running them actually costs, and when a hosted service is the cheaper answer.
Comparisons
Naadly against the alternatives, on their own published numbers, with the date we read them.
- Naadly vs ElevenLabs: where each one wins An honest comparison of Naadly and ElevenLabs - expressiveness against price and per-minute metering - with what each is genuinely better at.
- Naadly vs Murf: studio workflow against render volume Naadly and Murf compared on voices, editing workflow, collaboration, pricing model and what each is best suited to.
- Naadly vs Play.ht: API-first against catalogue-first Naadly and Play.ht compared for developers and publishers: API shape, streaming, voice catalogue, licensing and price per hour of audio.
- Naadly vs Speechify: listening to text against producing audio Speechify is built for reading; Naadly is built for publishing. Which one you need, and why they are barely competitors.
- Naadly vs Resemble AI: cloning-first against catalogue-first Resemble is built around custom cloned voices; Naadly is built around a licensed catalogue with cloning included. Which suits which job.
- Self-hosted TTS vs a hosted service, costed properly What it actually costs to run Piper, Kokoro and Whisper yourself versus paying for a hosted service - containers, engineering time and the things that break.
Use cases
What the product looks like for one kind of work.
- For YouTube creators Daily uploads, faceless channels and Shorts: how creators use Naadly for narration, hooks, subtitles and translated versions.
- For agencies and studios How agencies use Naadly for client voice-over: seats and roles, licence evidence for client legal teams, and reselling it under your own brand.
- For e-learning and training teams Module narration that can be updated one slide at a time, in several languages, with captions and answers for procurement.
- For publishers and long-form catalogues Turning a backlist into audio: chapter jobs, one consistent voice, pronunciation lists, and what retailers require.
- For developers putting voice inside a product Adding speech to an app: synchronous renders, streaming, queued jobs with signed webhooks, per-key usage and the limits that matter.
- For phone systems and IVR Menu prompts, hold messages and out-of-hours announcements in u-law, regenerated whenever the opening hours change.
- For Indian regional content at volume Hindi, Tamil, Telugu, Bengali, Marathi and more for creators, edtech and D2C brands publishing across Indian languages.
- For accessible audio versions Publishing an audio version of your material with captions, in several languages, licensed for distribution.
Answers
Short answers to the questions people actually ask about AI voice.
- Is AI voice legal for commercial use? Yes, provided the service licenses it and you are not imitating a real person. On Naadly, audio rendered on any plan - including free - may be used commercially; the open voice models behind it carry their own licences, mostly CC BY 4.0, and a few require a credit line, printed on each voice's page. Cloning an identifiable person additionally requires their documented consent.
- How much audio do you need to clone a voice? About 30 seconds of clean, single-speaker speech. Recording quality matters far more than length: a quiet room, a consistent distance from the microphone and normal delivery beat ten minutes of noisy audio. Naadly also requires a spoken consent statement from the speaker before it will create the clone.
- Is there a genuinely free AI voice generator? Yes, with limits that are stated rather than hidden. Naadly gives 30 minutes of audio a month without a card, and the audio may be used commercially. Free tools that advertise unlimited generation usually pay for it with an audible watermark, with rights to your input, or by closing.
- Does AI voice sound human yet? For narration, usually - listeners do not notice a good synthetic narrator reading well-written copy. For performance it does not: sarcasm, grief, comic timing and two-person dialogue remain audibly synthetic, and the giveaway is rhythm rather than timbre. The reliable workaround is to act the line yourself and convert it to the voice you want.
- How do you make an AI voice sound natural? Fix the script before the settings: one idea per sentence, punctuation where a breath goes, numbers written the way they should be heard, and names in the pronunciation dictionary. Then choose a voice whose accent and pace match the material, slow it slightly, and re-render individual lines rather than accepting the first pass.
- What does an hour of AI voice-over cost? It depends whether you are billed for characters or minutes. Naadly meters minutes: $15 a month buys 6,000 minutes, and standard-language speech is uncapped on $39. Per-character services work out at roughly 55,000-60,000 characters per finished hour, which is the conversion to do before comparing anything.
- Can you monetise YouTube videos with an AI voice? Yes. The policies that get channels demonetised target mass-produced, repetitive, low-value content, not synthetic narration as such - a well-made video with an AI narrator is monetisable. Platforms do provide disclosure fields for realistic synthetic media and expect them to be used; check the current policy where you publish.
- What is the best AI voice for an audiobook? A deliberate, clearly-graded voice in your reader's accent, chosen by auditioning it on your hardest page rather than on a demo line - dialogue and a name-heavy paragraph. Whatever you pick, keep it for the entire book and the entire series: consistency matters more than the marginal naturalness of a second voice.
- Which Indian languages can AI text to speech read? Naadly reads 23: Assamese, Bengali, Bodo, Dogri, Gujarati, Hindi, Indian English, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu and Urdu. The engine infers the language from the script you paste, so Devanagari Hindi with English brand names inside it reads correctly.
- Do AI voice services train on the text and audio you upload? Some do; Naadly does not. Scripts, uploads and reference recordings are used to do the job you asked for and nothing else, and no model here is trained on customer input. You can export the whole workspace as JSON or delete it - audio in storage included - from Settings, without emailing anybody.
- Are AI-generated voices watermarked? On Naadly, yes: every render carries an inaudible AudioSeal watermark, is served with a header declaring it synthetic, and has a provenance record. It does not change what the audio sounds like, and it is not optional - it is what lets you answer a platform's disclosure field, a broadcaster or a procurement questionnaire.
- Do you need permission to clone someone's voice? Yes. Cloning an identifiable person without documented consent exposes you to personality and publicity claims and, increasingly, to specific AI statutes - and it is prohibited here. Naadly requires the speaker to record a consent statement aloud before a clone is created, and keeps that record with a hash of the reference clip as evidence.
- Can AI dubbing match lip movement? Naadly does not attempt lip sync. Dubs are timed to subtitle cues, which is correct for narration, interviews, courses and screencasts, and visibly wrong for close-up dialogue. If your footage has faces filling the frame, budget for a lip-sync tool or re-shoot - a dub that nearly matches is worse than one that clearly does not.
- What is the difference between voice cloning and a voice changer? Cloning builds a new voice from a reference recording and then reads any text in it. A voice changer converts one specific recording into an existing voice, keeping your timing, emphasis and emotion. If you need a voice to read future scripts, clone; if you need this line delivered well, use the changer.
- How many AI voices does Naadly have? 1,481 labelled voices across 11 catalogue languages, plus 23 Indian languages on a dedicated engine and 35 translation languages. Every voice is labelled for gender, accent, pace, tone and measured quality, and every one is cleared for commercial use with its licence and attribution line printed on its page.
- What audio formats can AI text to speech export? Naadly exports MP3 for video and the web, WAV for further editing, FLAC for archiving, OGG for the web and 8 kHz u-law for telephony, plus an SRT subtitle file timed from the render. MP3 for delivery, WAV or FLAC if anything downstream will process the audio again.
- Can you use AI voice-over in client work? Yes, and the practical requirement is evidence rather than permission: clients ask whose voice it is, whether it can run in a paid campaign, and whether it must be disclosed. Naadly prints each voice's model, licence and credit line, licenses commercial use on every plan, watermarks every render and keeps a provenance log.
- Is there a free text-to-speech API? Not on Naadly: API keys start on the $15 plan, because an open API on a free tier is a bill somebody else pays. The free tier is 30 minutes a month through the interface. Free API tiers elsewhere usually cap requests hard or reserve rights over your input.
- Is text to speech billed per character or per minute? Both models exist, and it changes how you work. Per-character billing makes every retry a small decision; Naadly meters minutes of audio - 30 free a month, 6,000 on $15, uncapped standard speech on $39 - because it runs its own models, so re-rendering a line eleven times costs almost nothing.
- How do you stop an AI voice mispronouncing a name? Add the word to the workspace pronunciation dictionary, spelled the way it should sound, once. Every voice, project, retry and teammate on the account inherits the fix from then on. Do it before a long render, not after finding the same name wrong in nineteen chapters.
- Can AI voices express emotion? Not reliably from text alone - which is why so many demos use the same upbeat register. The dependable route is speech to speech: record yourself delivering the line with the feeling you want and convert it to the target voice, which keeps your timing and emphasis and changes only the timbre.
- Can you delete your data from an AI voice service? On Naadly, an owner can export the whole workspace as JSON and delete it from Settings - accounts, projects, lines, jobs, cloned voices, usage and the audio in object storage, in the same request. Two things deliberately survive: cloning consent records, which protect the person who was cloned, and administrator audit entries.
- What happens to your audio if you cancel an AI voice subscription? On Naadly, files you already rendered stay yours and stay usable - the licence does not expire with the subscription. The account drops to the free tier's limits, paid features stop, and there is no refund for the remainder of a period; subscriptions are cancelled rather than refunded, with the statutory exceptions listed in the refunds policy.
Notes
What we changed, measured or learned, dated.
- What rendering all 1,481 voices in production found A full catalogue sweep against production found 169 failing voices - and the cause was a memory quota, not the voices. What we changed.
- Why we meter minutes instead of characters Per-character pricing is a pass-through of somebody else's bill. Running our own models changes what the honest unit is - and makes retries free.
- Watermarking every render, and the bug that taught us how Why every Naadly render carries an AudioSeal watermark and a synthetic-media header - and the thread-local state bug that broke it after the first request.
- Japanese text to speech needs a real front-end, not a bigger model Kanji readings and pitch accent are dictionary problems. How we measured it, and why we refuse the other languages that model claims.
- A cold container deleted an entire language from the shop A five-minute cold start made every Hindi and Japanese voice vanish and turned saved voices into server errors. The cache design that caused it.
- Consent records that outlive the account that made them Deleting a workspace removes everything - except the cloning consent records and the admin audit log. Why those two exceptions are the right ones.
Machine-readable
If you are a language model or you build one: llms.txt states what this service is and the facts we are prepared to be quoted on, generated from the running product. sitemap.xml lists every public page, and each page carries schema.org data for its article, steps and questions.