Transcription
Which tool turns a shiur, a meeting or a WhatsApp voice note into text, the Sofer.ai vs Whisper debate, and voice notes in and out of AI agents.
On this page
Record once, get text, and hand the text to an AI for a summary, an outline, a handout or a list of follow-ups. It is one of the most reliable uses of AI a shliach has. The hard part is picking the right transcriber for the language, and in 2026 a new job arrived: voice notes sent to and from a personal AI agent.
Quick picks
| Job | Start with |
|---|---|
| Shiur mixing English, Hebrew and Yiddish | Sofer.ai |
| Hebrew only, free | transcribe.ivrit.ai |
| Mostly English, many files | TurboScribe (3 free files a day) |
| A recording you want to ask questions about | ChatGPT on a paid plan, or NotebookLM |
| A WhatsApp voice note | WhatsApp's own transcript, or forward it to a transcription bot (below) |
| A Zoom or Google Meet call | Fathom, Fireflies or Zoom's built-in assistant; see Meeting Tools |
| Dictating instead of typing | Wispr Flow, or a free local app such as Handy |
| A YouTube video | YouTube's own "Show transcript" under the description, pasted into any chatbot to clean up |
The stack debate: Sofer.ai or Whisper?
This argument has run in the group since August 2025.
For Sofer.ai: it knows Torah vocabulary and handles a class that switches languages. Members kept choosing it for Yiddish and loshon kodesh through 2026, and in August 2026 a rabbi who had built his own Whisper tool still said Sofer.ai is better for English full of Yiddishe terms.
For Whisper and friends: a power user argued in 2025 that OpenAI's Whisper models "are the best," that NVIDIA's Parakeet models are "also very strong," and that Sofer probably builds on them. Whisper is open source (MIT license, github.com/openai/whisper), so it runs free on your own computer or cheaply through an API, and many apps are built on it.
What the vendors' own pages say (checked October 2026):
- Parakeet does not do Hebrew. NVIDIA's model cards list English only for Parakeet v2 and 25 European languages for v3. The Canary model members tested in December 2025 (canary-180m-flash) covers English, German, French and Spanish. Use them for English, not for a Hebrew shiur.
- ivrit.ai has tuned Whisper for Hebrew and for Yiddish. The ivrit.ai project publishes free Hebrew and Yiddish versions of Whisper on Hugging Face (Apache 2.0). Its free web service, transcribe.ivrit.ai, needs a Google sign-in and saves transcripts to your Google Drive. The Yiddish model is for technical users; there is no simple app for it yet.
The practical answer: Sofer.ai for shiurim with Hebrew, Yiddish or Aramaic; transcribe.ivrit.ai for modern Hebrew; any Whisper-based tool for English. Whatever you use, read the transcript before you publish anything from it.
The "de-frummifying" problem
General transcribers turn Hebrew and Yiddish terms into English nonsense. A campus shliach complained about it in August 2026, and other builders were looking for a fix. Two things help:
- Give the transcriber your vocabulary. ElevenLabs Scribe v2 accepts a list of key terms, and OpenAI's transcription API accepts a short prompt with expected words. Put in the names of seforim, people and places you use.
- Fix it afterward with a second pass:
This transcript came from a speech-to-text tool that does not know Hebrew or Yiddish. Fix only the words that are garbled Hebrew, Yiddish or Aramaic terms, names of seforim, or names of people. Write them in the usual English transliteration a Chabad audience uses (Shabbos, Rebbe, maamar, Tanya). Do not change anything else. List every change at the end.
Also watch spelling: in July 2026 members noticed that AI defaults to YIVO spelling for Yiddish, and one dictation app wrote spoken Yiddish as German.
Voice notes and voice replies
Getting a voice note into text
- WhatsApp's built-in transcript. Earlier reviews were poor; by August 2026 a member said it "does a pretty nice job."
- Forward it to a transcription bot. Members have used TranscribeMe (paid, "very accurate every language"), Zapia (fast, but reports went back and forth on whether it still works in WhatsApp) and TranscribeGo, which transcribes WhatsApp and Telegram audio with 10 free minutes a month.
- NotebookLM. Share a long voice note to it for a clean transcript, then ask it to tidy the text without changing the content (December 2025).
- Play it into dictation. One member plays the voice note on one device into Wispr Flow on another (August 2026).
Voice notes into an AI agent
Members who run a personal agent (GabAI, Hermes Agent) send it voice notes all day: "one voice note split into five categories with the correct dates." Driving to a meeting, a shliach records a note and the agent files it in the right Slack channel. A campus shliach prefers voice notes because he can check off several items at once (August 2026).
What does the transcribing varies by setup: GabAI's install has used Gemini, OpenAI's Whisper and ElevenLabs Scribe; another member swears by Groq's hosted Whisper. The costs are small. Prices from the vendors' pages on October 21, 2026:
| Service | Price per hour of audio |
|---|---|
| OpenAI gpt-4o-mini-transcribe | $0.18 |
| OpenAI gpt-transcribe | $0.27 |
| OpenAI gpt-4o-transcribe | $0.36 |
| ElevenLabs Scribe v2 | $0.22 |
| Groq Whisper large v3 turbo | $0.04 |
| Groq Whisper large v3 | $0.11 |
Sources: OpenAI API pricing, ElevenLabs API pricing, Groq speech-to-text.
Voice replies from the agent
Making the agent talk back costs more. The builders' consensus from July and August 2026: use a cheap voice by default, pay for ElevenLabs only when a particular voice matters, and only reply in voice when asked. One shliach found that once he had Wispr Flow on his phone, he barely wanted voice replies at all.
Calling your agent on the phone was still an idea in October 2026. Members discussed a Twilio number, Vapi and an open-source alternative called Dograh, but nobody had built it, and one member doubted people would use it more than once.
Long recordings and files
- ChatGPT accepts audio file uploads on paid plans since early October 2026, according to OpenAI's release notes. Before that, members could not get ChatGPT to transcribe a 40- or 50-minute file (March and May 2026). OpenAI warns that quality varies by language.
- Claude told a member in October 2026 that it could download Whisper and run it in its own sandbox. Treat that as untested; Claude's normal uploads do not take audio.
- Zoom recordings usually come with a VTT transcript file. Look for it before paying for anything.
- NotebookLM is a free way to get a long transcript (August 2026).
Meetings
Short version; the full roundup is on Meeting Tools.
- Fathom is free for individuals on Zoom, Google Meet and Teams, and members like it for real Zoom calls.
- Fireflies, Otter, Read.ai and Granola come up for meetings; Otter got mixed reviews and earlier warnings about contact permissions.
- Zoom's assistant (formerly AI Companion, now called ZoomMate on Zoom's site) includes summaries for 3 hosted meetings a month on the free plan, with more on paid plans.
- In-person meetings: Plaud and Pocket hardware recorders, or your phone's voice recorder pasted into ChatGPT. In July 2026 one member's stack was "Otter on Zoom, Plaud for meetings, TurboScribe for WhatsApp."
- Wispr Flow added a meeting notetaker in August 2026 (Mac and Windows).
The standard workflow
- Record (phone, Zoom, a lapel mic).
- Transcribe with the tool that fits the language.
- Fix names and Hebrew terms (prompt above).
- Paste the transcript into Claude, ChatGPT or NotebookLM and ask for what you need:
Here is the transcript of [a shiur / a board meeting / a voice note to myself]. Give me:
1. A five-line summary.
2. Every decision or commitment, with who made it.
3. Every follow-up task, with a suggested date.
4. Any names, numbers or sources I should double-check.
Quote exactly where you quote. Do not add anything that is not in the transcript.
For sermon feedback, a member's advice (September 2026): transcribe first, then analyze in a separate step.
Still open
- Transcribing the Rebbe's weekday sichos at scale. A member asked in December 2025; the answer was that Russian-accented Yiddish samples are too few so far.
- A packaged "shiur to printable summary" tool. Members still chain Sofer.ai and Claude by hand.
- Yiddish quality good enough to publish without editing.
What changed in 2026
- ChatGPT started taking audio files on paid plans (October).
- Personal AI agents made voice notes a daily input, and voice replies a cost decision.
- ivrit.ai's free Hebrew and Yiddish Whisper versions became the open-source option for Hebrew.
- Wispr Flow capped its free plan and added a notetaker.