# Best AI voice generator in 2026: text to speech, transcription, cloning and live voice compared > To read text aloud for free, start with ElevenLabs' free plan or Gemini's text to speech on Google's free API tier. To turn speech into text privately, run Whisper on your own computer. To clone a voice, expect a consent step: ElevenLabs' professional clones take your own voice only. Prices and plans below are as listed on 6 November 2026. Mindket editors · Published 2026-11-05 · Updated 2026-11-06 · https://mindket.com/learn/ai-tools/best-ai-voice-generator ## In short - Name the voice job first, because reading aloud, transcribing, cloning and live talk are sold and licensed apart. - Get the speaker's recorded agreement before you try to clone anyone's voice. - Narrate a monetized channel only on a plan that lists a commercial license. - Keep recordings you may not share on your own machine. ## Voice platforms compared: free plans, prices and limits Eight platforms cover most voice jobs. ElevenLabs, OpenAI, Google and Microsoft sell finished voices and transcripts in the cloud. Deepgram and AssemblyAI sell speech recognition and text to speech to people building apps, Deepgram's recognition by the minute and AssemblyAI's by the hour of audio. Whisper and Kokoro are free models you run yourself. Read the free plan column first. Most of these let you hear your own script, or transcribe your own file, before you pay anything. Each vendor counts in its own unit (credits, tokens, minutes or characters), so a price only means something next to the length of your audio. These prices change often. | Platform | Made for | Free plan | Entry paid plan, as its page states it | Limits that bite | Commercial use | API | Runs on your computer | | --- | --- | --- | --- | --- | --- | --- | --- | | ElevenLabs | Reading text aloud, voice cloning, transcription (Scribe), voice changer and voice agents | 10,000 credits a month, about 10 minutes of text to speech | Starter: $6 per month, $1 the first month, 30,000 credits | About 30 minutes of speech a month on Starter. Professional clones start at Creator | Starter and up include a commercial license. The free plan lists none | Yes | No | | OpenAI API | Text to speech, transcription and live voice agents inside your own app | None listed on its pricing page | Pay as you go: gpt-transcribe $0.0045 a minute, GPT-Live-1 $0.05 a minute, gpt-4o-mini-tts $12 per 1M audio tokens | Transcription files up to 25 MB. tts-1, tts-1-hd and gpt-4o-mini-tts shut down on 6 January 2027. Voices are tuned for English | Its usage policies require telling listeners the voice is AI-generated | Yes | No | | Google Gemini API | Text to speech with one or two speakers, voice replication (cloning, paid tier only), and live voice through the Live API | Gemini 3.8 Flash TTS and Gemini 3.8 Live are free of charge on the free tier | Paid tier: Gemini 3.8 Flash TTS $9.00 per 1M audio tokens out until 31 December 2026, $18.00 from 1 January 2027 | Two speakers per request at most. Free tier inputs and outputs may be read by human reviewers | "Google won't claim ownership" of the output, its API terms say | Yes | No | | Microsoft Azure Speech | Speech to text, neural text to speech and personal voice in business apps | Free (F0): 5 audio hours of speech to text and 0.5 million characters of neural text to speech a month | Not stated on its page | Personal voice and professional voice need Microsoft's approval first | Not stated on its page | Yes | No | | Deepgram | Speech to text, text to speech and voice agents for developers | $200 of credit, no card needed, no expiry | Pay as you go: Nova-3 $0.0043 a minute for recordings, $0.0077 a minute live, Aura-2 $0.030 per 1,000 characters | Up to 150 live speech to text streams at once on pay as you go | Not stated on its page | Yes | No | | AssemblyAI | Speech to text for recordings and live audio, and text to speech (AssemblyAI-TTS-1) for voice agents | $50 of credit, no card needed | Pay as you go: Universal-3.5 Pro $0.21 an hour of audio, live streaming from $0.15 an hour, AssemblyAI-TTS-1 $15 per 1M characters | Free tier: 5 new live streams a minute. Its Voice Agent API is being retired | Not stated on its page | Yes | No | | Whisper, open source from OpenAI | Speech to text on your own machine | Free: code and weights under the MIT License | None | The turbo model needs about 6 GB of video memory. Accuracy varies widely by language | MIT License | Python library | Yes | | Kokoro-82M, open weights | Reading text aloud on your own machine | Free: Apache 2.0 weights | None | 8 languages and 54 preset voices. Its page describes no cloning, and says non-English support may be thin | Apache 2.0, which its page says suits production use | Python library | Yes | > Google Cloud's Speech-to-Text pricing page showed no figures when we read it, so Google is here through the Gemini API. ## Pick by what you need the voice for **Text to speech** reads your words aloud. **Speech to text** writes down what someone said. **Voice cloning** builds a copy of one real person's voice, and **speech to speech** swaps the voice on a recording you already have. **Live voice** is a spoken conversation with an app, with no wait for a file. Start from the job in front of you, then look up its row. - **Free narration to try a script:** ElevenLabs' free plan, or Gemini's Flash TTS on Google's free tier. Google suggests trying its voices in AI Studio before you build anything, and that needs no code. - **Narration for a channel you monetize:** ElevenLabs Starter or above, the first plan with a commercial license. With OpenAI's voices, also tell listeners the voice is AI-generated. - **Transcripts of meetings or interviews:** Whisper on your own computer costs nothing per minute. In the cloud, gpt-transcribe, Deepgram Nova-3 and AssemblyAI charge by the minute or hour, and the last two start you with credit. - **Cloning your own voice:** ElevenLabs, from Starter for an instant clone and from Creator for a professional one, or the Gemini API on its paid tier. - **Changing the voice on a recording:** ElevenLabs Voice Changer, in pieces of up to 5 minutes, across 29 languages with its multilingual model. - **A live voice agent in your app:** OpenAI's guide recommends GPT-Live, a model that listens while it speaks, for a new conversational app. Gemini Live costs nothing on Google's free tier. Deepgram sells a Voice Agent API by the minute. - **Private audio, or no internet:** Whisper for transcripts and Kokoro for speech, both on your own machine. A local text model run through [Ollama](https://mindket.com/learn/local-ai/ollama-tutorial) can write the script Kokoro reads. - **Chinese voice or text:** Gemini's text to speech lists Chinese in simplified and traditional script, plus Cantonese. Kokoro has 8 Mandarin voices. For transcripts, OpenAI's gpt-transcribe accepts regional Chinese language codes. - **An API for your own app:** every cloud platform in the table sells one. OpenAI's is the one our [API quickstart](https://mindket.com/learn/openai-api/openai-api-quickstart) sets up, key first. - **A spoken summary of your documents, not a voiceover:** that is what an Audio Overview in [NotebookLM](https://mindket.com/learn/ai-assistants/notebooklm-tutorial) makes from your files. **How many languages each voice model lists** Each model's own page, read on 6 November 2026. "Over 130" and "90+" are drawn at 130 and 90. - Gemini 3.8 Flash TTS (text to speech): 130+ - Whisper, as whisper-1 in OpenAI's API (speech to text): 98 - Azure personal voice (cloned voice): 91 - ElevenLabs Eleven v4 (text to speech): 90+ - ElevenLabs Flash v2.5 (text to speech): 32 - ElevenLabs Voice Changer, multilingual (speech to speech): 29 - Kokoro-82M v1.0 (text to speech): 8 Sources: [Google: Gemini API speech generation](https://aistudio.google.com/docs/speech-generation) (read 2026-11-06), [OpenAI: Speech to text guide](https://developers.openai.com/api/docs/guides/speech-to-text) (read 2026-11-06), [Microsoft: What is personal voice?](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/personal-voice-overview) (read 2026-11-06), [ElevenLabs: Models](https://elevenlabs.io/docs/overview/models) (read 2026-11-06), [ElevenLabs: Voice changer](https://elevenlabs.io/docs/overview/capabilities/voice-changer) (read 2026-11-06), [Hugging Face: hexgrad/Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M) (read 2026-11-06) ## What each vendor asks before it clones a voice A voice clone is a model of one person's voice, built from recordings of them. None of the six services below makes one without a consent step, and three of them admit only customers they have approved. OpenAI and Google go furthest on wording. The speaker reads a fixed sentence aloud, and OpenAI says any change to the script makes the attempt fail. These are the English versions from their docs: ````text OpenAI: I am the owner of this voice and I consent to OpenAI using this voice to create a synthetic voice model. Google: I am the owner of this voice, and I consent to Google using this voice to create a synthetic voice model. ```` | Service | Audio it needs | The consent step | Who can use it | | --- | --- | --- | --- | | ElevenLabs Instant Voice Clone | About 1 to 2 minutes of clean audio is recommended. Eleven v4 can capture a voice from 10 seconds, the Eleven v4 announcement says | You confirm that you have the right and consent to clone the voice before you save it | Starter plan and above | | ElevenLabs Professional Voice Clone | 30 minutes at the least, 2 to 3 hours recommended | Your own voice only, even if someone else agrees. You read verification lines aloud to prove it | Creator plan and above | | OpenAI custom voices | A 10 to 60 second recording with several full sentences | The same speaker reads the consent phrase above, word for word | Eligible customers, through OpenAI's sales team | | Google Chirp 3 Instant Custom Voice | Up to 10 seconds of single-channel audio | The speaker records Google's consent statement above, and no custom script is allowed | Allow-listed users, through Google's sales team | | Gemini API voice replication | A 10 to 30 second clip of clean, natural speech, plus a consent recording, both from the same adult speaker | The same speaker recites Google's consent statement aloud, in one of 30 supported locales | Paid tier (Tier 1 or higher) Gemini API projects, with no allowlist | | Microsoft Azure personal voice | A 5 to 90 second speech sample | A recorded statement from the speaker that the Azure customer will create and use their voice | Approved customers and use cases, through an intake form | > Open models skip all of this. Chatterbox, Resemble AI's open model, cloned a voice from a reference clip and stamped every file with its Perth watermark. Its page says it was retired in August 2026. With an open model, getting consent is your job alone. ## Make your first voiceover free in three steps ElevenLabs' free plan lets you hear your own script in several voices in the browser, with no code. The menu names below are from its text to speech guide. Every generation spends credits. The first listen counts too. Regenerating the same text with the same voice and model is free twice, within two hours, as long as you have not refreshed the page. 1. **Open Text to Speech.** Sign up at [elevenlabs.io](https://elevenlabs.io), select Text to Speech in the sidebar and paste your script into the input box. One generation takes up to 10,000 characters with Eleven v4 and 5,000 with the other models. 2. **Choose a voice and a model.** Pick a voice from your Voices at the bottom left, then a model in the model selector. With Eleven v4, a tag such as [laughs] in the script directs the delivery. 3. **Generate and download.** Select Generate speech, listen, then press the download button at the bottom right. For WAV or another format, open the history tab in the panel on the right and click the download icon. It offers MP3 at 128 kbps or WAV, and Advanced adds MP3 at 192 or 256 kbps, M4A and FLAC. ## Transcribe offline with Whisper Whisper is the speech recognition model OpenAI released under the MIT License. It runs on your own computer, so the recording never leaves it. No account is needed. You need Python and ffmpeg, the command-line audio tool (`choco install ffmpeg` or `scoop install ffmpeg` on Windows, `brew install ffmpeg` on a Mac). These two lines from its README install it, then transcribe three files with the turbo model: ````bash pip install -U openai-whisper whisper audio.flac audio.mp3 audio.wav --model turbo ```` > Turbo needs about 6 GB of video memory and is not trained to translate, so pick another size if you want English out of a foreign recording. The tiny model needs about 1 GB and runs about ten times as fast as large. Your graphics card decides which you can run, as the [hardware lesson](https://mindket.com/learn/local-ai/local-llm-hardware) explains. ## How ElevenLabs and OpenAI's developer team say to direct a voice On 28 September 2026, the day Eleven v4 came out, ElevenLabs [posted on X](https://x.com/ElevenLabs/status/2104572133636731068) that you can direct its performance with inline tags. A tag is a short instruction in square brackets inside the script. You can set the delivery, emotion, pacing, reactions, sound effects or style that way, or just describe how a line should sound. Try it in step 2: generate one line with a tag and once without, and keep whichever reads your line the way you meant it. OpenAI's developer account [announced on 10 September 2026](https://x.com/OpenAIDevs/status/2098099269551149398) that GPT-Live-1 had reached the API, for voice agents that listen while they speak. That matters only for live voice. A recorded voiceover never gets interrupted. A caller does. When you try an agent, cut in mid-sentence and see whether it stops to listen. ElevenLabs' own YouTube channel has the launch video for both new voice models. > Creators can specify delivery, emotion, pacing, reactions, SFX, style, or simply describe how a line should be delivered > > ElevenLabs, the company that makes Eleven v4, on its own account, [X](https://x.com/ElevenLabs/status/2104572133636731068) (read 2026-11-05) [![ElevenLabs (@ElevenLabs) on X, 28 September 2026: Direct the performance of Eleven v4 with inline tags. Creators can specify delivery, emotion, pacing, reactions, SFX, style, or simply describe how a line should be delivered.](https://mindket.com/blog/shots/x-elevenlabs-v4-inline-tags.webp?v=e7ccd715b6c5)](https://x.com/ElevenLabs/status/2104572133636731068) ElevenLabs on X, 28 September 2026 [![OpenAI Developers (@OpenAIDevs) on X, 10 September 2026: GPT-Live-1 is now available in the API. Bring ChatGPT's natural back-and-forth to your app, with voice agents that listen while they speak.](https://mindket.com/blog/shots/x-openaidevs-gpt-live-1-api.webp?v=361ead698a11)](https://x.com/OpenAIDevs/status/2098099269551149398) OpenAI Developers on X, 10 September 2026 [Introducing Eleven v4 and Eleven v4 Turbo](https://www.youtube.com/watch?v=th_tXR2QQ6U) ElevenLabs' own YouTube channel introduces Eleven v4 and Eleven v4 Turbo. > Screenshots of X's own embeds of the posts, taken on 5 November 2026, with the time in UTC. Select one to open the post on X. ## Common mistakes - Judging a voice by the vendor's demo reel. Paste your own script, with your product names and numbers, since that is the text it has to get right. - Starting a new app on tts-1 or gpt-4o-mini-tts. They shut down on 6 January 2027, and OpenAI points new work to gpt-realtime-2.1-mini through the Realtime API, a different workflow. - Sending a long recording to OpenAI's transcription endpoint in one piece. Anything over 25 MB is refused. Split it first. - Putting confidential calls through the Gemini API's free tier. Google says human reviewers may read free tier inputs and outputs, so use the paid tier or a local model for those. ## Try it yourself - Paste the same short paragraph into ElevenLabs and into Gemini text to speech in AI Studio, and note which one says your product names right. - Transcribe a two-minute voice memo with Whisper's tiny model, then with turbo, and mark the words each one gets wrong. - Pick the table row you would use for your next project, and write down the one limit that could stop you. ## Check yourself 1. You will narrate videos on a channel you monetize with ElevenLabs. Which plan is the first to list a commercial license? A. Free B. Starter C. Professional Voice Clone D. Enterprise only Answer: B. ElevenLabs' pricing page lists a commercial license from Starter up, and none on the free plan. 2. A friend agrees to let you make an ElevenLabs Professional Voice Clone of their voice. What happens? A. It works after they read the verification lines B. It works on the Creator plan or above C. It is refused, because a professional clone must be your own voice D. It needs a 10-second sample Answer: C. ElevenLabs' docs say you can only make a Professional Voice Clone of your own voice, even with the other person's consent. 3. Which option transcribes a confidential recording without sending it anywhere? A. Whisper on your own computer B. The Gemini API's free tier C. AssemblyAI's sign-up credit Answer: A. Whisper runs on your own machine, while every cloud option receives the audio. 4. You are starting a new app that reads text aloud with OpenAI. Why not build it on gpt-4o-mini-tts? A. It speaks only Chinese B. It needs a consent recording C. It cannot make MP3 files D. It shuts down on 6 January 2027 Answer: D. OpenAI's guide says tts-1, tts-1-hd and gpt-4o-mini-tts shut down on 6 January 2027 and points new work to gpt-realtime-2.1-mini. ## Frequently asked questions ### How much does an AI voice generator cost? You can start for nothing. ElevenLabs, the Gemini API, Azure, Deepgram and AssemblyAI each have a free allowance or sign-up credit, and Whisper and Kokoro are free to run. Paid plans count credits, minutes or tokens, so price them against the length of your own audio. ### Can I use an AI voice on YouTube and make money from it? Yes, as long as the plan you make it on allows commercial use, so check the plan, not the voice. On ElevenLabs every paid tier qualifies. If you build OpenAI's voices into an app, its policies ask for a clear disclosure to listeners. ### What is the best free AI voice generator? The one whose limit you can live with. A short script fits in ElevenLabs' monthly credits. Long audiobooks or daily jobs suit Kokoro, which has no cap and needs no account but offers only 54 preset voices. ### Is it legal to clone someone else's voice with AI? Not without their consent on any platform here, and the law where you live may add more. ElevenLabs' professional clones go further and accept only your own voice, even when the other person agrees. ### Which AI is best for transcribing audio to text? Deepgram or AssemblyAI if you need live captions in an app, since both sell streaming speech to text. To label who said what, OpenAI's guide names gpt-4o-transcribe-diarize, due to shut down on 26 February 2027. Private files belong with Whisper. ## Sources - [ElevenLabs: Pricing](https://elevenlabs.io/pricing) (read 2026-11-06) - [ElevenLabs: Models](https://elevenlabs.io/docs/overview/models) (read 2026-11-06) - [ElevenLabs: Text to Speech guide](https://elevenlabs.io/docs/product-guides/playground/text-to-speech) (read 2026-11-06) - [ElevenLabs: Instant Voice Cloning](https://elevenlabs.io/docs/product-guides/voices/voice-cloning/instant-voice-cloning) (read 2026-11-06) - [ElevenLabs: Professional Voice Cloning](https://elevenlabs.io/docs/product-guides/voices/voice-cloning/professional-voice-cloning) (read 2026-11-06) - [ElevenLabs: Voice changer](https://elevenlabs.io/docs/overview/capabilities/voice-changer) (read 2026-11-06) - [ElevenLabs: Introducing Eleven v4](https://elevenlabs.io/blog/eleven-v4) (read 2026-11-06) - [OpenAI: API pricing](https://developers.openai.com/api/docs/pricing) (read 2026-11-06) - [OpenAI: Audio and speech guide](https://developers.openai.com/api/docs/guides/audio) (read 2026-11-06) - [OpenAI: Text to speech guide](https://developers.openai.com/api/docs/guides/text-to-speech) (read 2026-11-06) - [OpenAI: Speech to text guide](https://developers.openai.com/api/docs/guides/speech-to-text) (read 2026-11-06) - [OpenAI: Custom voices guide](https://developers.openai.com/api/docs/guides/custom-voices) (read 2026-11-06) - [Google: Gemini API speech generation](https://aistudio.google.com/docs/speech-generation) (read 2026-11-06) - [Google: Gemini API pricing](https://aistudio.google.com/docs/pricing) (read 2026-11-06) - [Google: Gemini API additional terms of service](https://aistudio.google.com/docs/terms) (read 2026-11-06) - [Google: Gemini API voice replication](https://aistudio.google.com/docs/voice-replication) (read 2026-11-06) - [Google Cloud: Chirp 3 Instant Custom Voice](https://docs.cloud.google.com/text-to-speech/docs/chirp3-instant-custom-voice) (read 2026-11-06) - [Microsoft: Azure Speech pricing](https://azure.microsoft.com/en-us/pricing/details/cognitive-services/speech-services/) (read 2026-11-06) - [Microsoft: What is personal voice?](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/personal-voice-overview) (read 2026-11-06) - [Deepgram: Pricing](https://deepgram.com/pricing) (read 2026-11-06) - [AssemblyAI: Pricing](https://www.assemblyai.com/pricing) (read 2026-11-06) - [GitHub: openai/whisper](https://github.com/openai/whisper) (read 2026-11-06) - [Hugging Face: hexgrad/Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M) (read 2026-11-06) - [Hugging Face: Kokoro-82M voices](https://huggingface.co/hexgrad/Kokoro-82M/blob/main/VOICES.md) (read 2026-11-06) - [GitHub: resemble-ai/chatterbox](https://github.com/resemble-ai/chatterbox) (read 2026-11-06) - [ElevenLabs (@ElevenLabs) on X: Eleven v4 inline tags](https://x.com/ElevenLabs/status/2104572133636731068) (read 2026-11-05) - [OpenAI Developers (@OpenAIDevs) on X: GPT-Live-1 in the API](https://x.com/OpenAIDevs/status/2098099269551149398) (read 2026-11-05) - [ElevenLabs on YouTube: Introducing Eleven v4 and Eleven v4 Turbo](https://www.youtube.com/watch?v=th_tXR2QQ6U) (read 2026-11-05) [Browse automations](https://mindket.com/automations)