ElevenLabs
Text-to-speech, voice cloning and transcription; the group's pick when a voice has to sound like a person, Hebrew included.
On this page
ElevenLabs (elevenlabs.io) turns text into speech, clones voices, and transcribes audio. In the group it is the usual answer when someone needs narration that does not sound robotic: a shiur promo, a voicemail greeting, a children's story read in your own voice, or an AI agent that replies with a voice note.
What it costs
Checked on the pricing page on October 21, 2026. Every plan runs on credits shared across all ElevenLabs tools. For text-to-speech, one character costs one credit.
| Plan | Price | Credits a month | What it adds |
|---|---|---|---|
| Free | $0 | 10,000 (about 10 minutes of speech) | No commercial license |
| Starter | $6/month | 30,000 | Commercial license, instant voice cloning |
| Creator | $22/month ($11 the first month) | 121,000 | Professional voice cloning |
| Pro | $99/month | 600,000 | Higher audio quality through the API |
Scale and Business plans exist for teams. The free plan has no commercial license, so a voice used in a paid ad or a fundraising video needs Starter or above.
Nonprofits can apply to the ElevenLabs Impact Program, which gives approved nonprofits free access. ElevenLabs decides who qualifies, so apply and see.
In July 2026 a member who used it for agent voice replies said moderate use needed the Creator plan.
Hebrew
ElevenLabs' own model list includes Hebrew in Eleven v3 and in the newer Eleven v4 family, which launched at the end of September 2026. The older Multilingual v2 and Flash v2.5 models do not list Hebrew, so choose v3 or v4 for anything in Hebrew.
How the group got here: in November 2025 a member reported that v3 "can pronounce a ches reliably," and in March 2026 another said v3 finally handled Hebrew where earlier voices sounded Arabic. Nobody in the group has reported on v4 yet.
Getting Ashkenazi pronunciation
By default the voices read Hebrew the Israeli way. Members use two fixes:
- Give the voice Hebrew with nekudos instead of bare letters. In October 2026 a member making daily Chitas videos found that converting the Hebrew words in the script to vowelized Hebrew before narrating fixed most mistakes, and he added a list of Ashkenazi pronunciations to the prompt.
- Spell the words in English letters the way they should sound:
Khabad,samayakh,Shabbos,Sukkos.
To prepare a script, paste it into any chatbot with this prompt:
Below is a script in English with Hebrew and Yiddish words mixed in. A text-to-speech voice will read it aloud. Rewrite only the Hebrew and Yiddish words so the voice says them in Ashkenazi (Chabad) pronunciation. Write them in English letters, with "kh" for ches and chof, "s" for a tav without a dagesh, and "o" for a kamatz (Shabbos, Sukkos, Khabad, samayakh). Do not change any English words. Return the full script, then a list of every word you changed.
Listen to the whole result before you publish. The same spelling trick works in Suno.
What shluchim use it for
- Narration from a script, in a stock voice or your own cloned voice. Recommended for narration in your own voice in August 2026.
- Fixing two words in a recording without re-recording (Descript does this too), March 2026.
- A professional voicemail greeting: "Fiverr or ElevenLabs," June 2026.
- Listening to a long document. The free ElevenLabs Reader app reads PDFs and articles aloud. In August 2026 members suggested it for hearing a speech in a natural voice, and ElevenLabs itself when you need an MP3 file.
- Voice replies from an AI agent. See below.
- Transcription. Scribe v2 is ElevenLabs' speech-to-text model, priced at $0.22 per hour of audio on the API pricing page, with speaker labels and a list of key terms you can feed it. The GabAI media skill uses Scribe for audio once an ElevenLabs key is connected.
Voice replies from agents
Builders in the GabAI and Hermes groups compared voices in mid-2026. OpenAI's text-to-speech costs a small fraction of ElevenLabs and is decent; ElevenLabs "sounds like a real person" while with OpenAI "you can tell it's AI." The rule several of them settled on:
- Use the cheap voice by default.
- Pay for ElevenLabs only when a particular voice matters.
- Send a voice reply only when asked ("answer me back with voice"), so you are not paying to speak every message.
In October 2026 a member suggested a Twilio phone number plus an ElevenLabs voice agent for a phone and WhatsApp Q&A line. Nobody had built it yet. See Voice notes and voice replies for the incoming side.
Watch out for
- Credits go faster than expected. Ten minutes of speech is about 10,000 characters, the whole free month.
- Clone a voice only with that person's permission.
- Hebrew still needs a human ear. Even with nekudos, names and rare words slip.
Alternatives
- OpenAI text-to-speech, or ChatGPT and Gemini "read aloud": cheaper or free, more obviously synthetic. In August 2026 a member found ChatGPT and Gemini read-aloud more conversational than Claude's.
- Speechify: for reading documents to yourself, with a free plan.
- Descript: edit a recording by editing its transcript, with its own AI voices.
- Cartesia: another text-to-speech provider, mentioned but not reviewed in the group.
What changed in 2026
- Hebrew moved from a workaround to a listed language in v3, and v4 (September 2026) lists it too.
- Scribe v1 is deprecated; Scribe v2 is the current transcription model.
- The group's main use shifted from narration to voice replies for personal AI agents.