# Thoth: Private AI Meeting Scribe for Mac, iPhone, and iPad > Record both sides of any meeting. Transcribe, identify speakers, and summarize locally on your device. No cloud, no data ever leaving your machine. Homepage: https://thoth-app.com/ Mac App Store: https://apps.apple.com/app/thoth-your-private-ai-scribe/id6756965313?mt=12 iPhone and iPad (open beta): https://testflight.apple.com/join/hu2xjnN7 Developer: Matthieu Veinhard (solo, no funding, no team) Version: 1.7.0 (macOS). iPhone and iPad in open beta on TestFlight. Requires: macOS Tahoe (macOS 26) or later. Apple Silicon recommended. Intel Macs supported for recording and transcription. --- ## What Thoth Does Thoth is a native SwiftUI app for Mac, iPhone, and iPad that records meetings privately. On Mac it captures your microphone and the system audio (Zoom, Teams, Meet, or any app) on two separate channels simultaneously, transcribes everything on-device using WhisperKit and CoreML, detects individual speakers, and generates AI summaries. Nothing leaves your machine unless you explicitly choose to use a cloud AI key for summaries, and even then on-device anonymization can redact personal data first. Key distinction from cloud competitors: Thoth does not join your meetings as a bot, does not upload your audio to any server, and does not require an account or internet connection. One Pro purchase unlocks every platform. --- ## Features ### Recording On Mac, captures microphone and system audio independently. No virtual audio drivers, no screen capture, no bot joining the call. Mic channel is always Speaker 1 (you). System audio captures all remote participants. Both channels are recorded, transcribed, and diarized separately, making speaker attribution deterministic rather than estimated. On iPhone and iPad, Thoth records microphone audio (system audio capture from other apps is not possible on iOS due to platform restrictions). ### Live Transcript A floating panel displays your transcript word by word as you speak. Two models run in parallel: a fast streaming model (Parakeet EOU, ~160 ms latency) for instant preview, and a quality batch model for the final text. All on-device. ### Speaker Detection (Diarization, Pro) An on-device engine using Pyannote CoreML detects individual speakers and color-codes the transcript automatically. Benchmarked at 7.72 seconds to diarize a 42-minute recording with 2 speakers. Supports up to 8 speakers. Works offline. ### On-Device Anonymization Before any transcript is sent to an optional cloud AI, Thoth finds and masks names, places, phone numbers, and emails entirely on-device, swapping them for tokens. The mapping is reversible locally, so the summary reads naturally while the cloud provider never sees the original personal data. ### AI Summary Extracts notes, action items, and key decisions in one click. Choice of local on-device models or your own cloud API key (OpenAI, Anthropic, Google). Audio never leaves your machine. If using cloud AI, only the text transcript is sent directly from your Mac to your chosen provider with your own key, and anonymization can redact it first. ### Saved Custom Prompts Write and save your own AI instructions, then rerun any saved prompt on any recording. Build a personal library of summary styles, extraction templates, or question sets. ### On-Device AI Models Several local models available, ranging from small and fast to high quality (Ministral 3B through Gemma 3 12B). No API key required, no internet needed. ### Live Activity and Dynamic Island (iPhone) An active recording surfaces in the iPhone Live Activity and Dynamic Island with a running timer and interactive controls to stop or pause without opening the app. ### Device Transfer Move recordings directly between iPhone and Mac over the local network or AirDrop. No cloud round-trip. ### Export Formats Free: WAV audio, TXT transcript. Pro: M4A, AAC, Markdown, RTF, JSON (with timestamps and speaker colors), PDF. Direct sharing via Mail, Messages, AirDrop. ### Language Support Whisper models support 99 languages with auto-detection. Parakeet TDT v3 supports 25 European languages and Japanese/Chinese for re-transcription (batch mode). Note: Parakeet has no language conditioning mechanism and will translate non-English audio to English on ambiguous speech (see blog post on language drift). The app interface is fully localized in English and French. ### Menu Bar Mode (Mac) Thoth can run as a compact always-on-top menu bar item. One-click recording start/stop without switching windows. --- ## Transcription Models and Benchmarks All benchmarks run on an Apple M2 MacBook Pro, 42-minute recording. ### Re-transcription (batch) | Model | Speed | WER (accented EN) | WER (clean EN) | Languages | |---|---|---|---|---| | Whisper Large V3 Turbo | 12.7x real-time | 32.3% | 7.8% | 99 | | Parakeet TDT V3 (Pro) | 180x real-time | 38.9% | 8.7% | 25 | | Whisper Small | 17.7x real-time | 45.9% | 9.2% | 99 | | Whisper Base | 59.6x real-time | 51.1% | 9.2% | 99 | Whisper Large V3 Turbo wins on accuracy across all scripts. Parakeet is 14x faster than Whisper Large V3 Turbo on the same file with near-identical WER on clean speech, but falls behind on accented speech. ### Live transcription | Engine | WER | Latency | |---|---|---| | Parakeet EOU 120M (Pro) | 38.4% | ~160 ms | | Parakeet Sliding Window (Pro) | 56.8% | ~11 s | | WhisperKit Base+Small | 65.9% | ~12 s | ### Diarization 7.72 seconds to diarize 42 minutes of audio with 2 speakers. Up to 8 speakers. Fully on-device using Pyannote CoreML. ### AI Summary Quality (local vs cloud) Tested on a real French-language interview transcript. Scored by Claude Opus across 6 criteria. | Criterion | Local (Qwen 7B) | Cloud (Claude Sonnet BYOK) | |---|---|---| | Factual accuracy | 7/10 | 9.5/10 | | Completeness | 5/10 | 9/10 | | Decision capture | 2/10 | 8.5/10 | | Action items | 5/10 | 8/10 | | Quote selection | 4/10 | 8.5/10 | | Language quality | 7/10 | 9/10 | | Overall | ~5/10 | ~8.7/10 | | Privacy | Zero data leaves | Text sent to provider | | Cost | Free | ~$0.01/hour | | Internet required | No | Yes | --- ## Privacy Architecture - Audio transcription: WhisperKit + CoreML, runs entirely on-device. Audio never leaves the device. - Speaker detection: Pyannote CoreML, fully on-device. - On-device AI summaries: local models (Ministral 3B, Gemma 3 12B, and others), no internet needed. - On-device anonymization: names, places, phone numbers, and emails are detected and masked locally before any optional cloud AI sees the transcript. - Cloud AI (optional, BYOK): only the text transcript is sent, directly from the user's device to their chosen provider (OpenAI, Anthropic, Google) using the user's own API key. Thoth never receives or stores the key or the transcript. - API keys stored in Apple Keychain, never in plaintext. - No account required. No data collection by Thoth. --- ## Pricing ### Free - $0 - Unlimited recordings - 30 min per mic recording, 15 min per system audio recording - 10 AI enhancements per month (local or cloud) - WAV audio export, TXT transcript export ### Pro - $9.99/month - $49.99/year - $99.99 lifetime (one-time purchase, also available at EUR 99.99) - Unlimited recording duration - System audio and mixed recording - Speaker identification (diarization) - All export formats: M4A, AAC, Markdown, RTF, JSON, PDF - Unlimited AI enhancements (local or cloud with own API key) - Large transcription model (Whisper Large V3 Turbo) - Parakeet TDT v3 and Parakeet EOU models - Remove branding from exports and shares One Pro purchase unlocks every platform. Free trial included with Pro subscription. Note: Cloud AI features (OpenAI, Anthropic, Google) may not be available in all countries due to local regulations. On-device AI is available everywhere. --- ## Comparison with Cloud Competitors | | Thoth | Otter | Fireflies | Granola | |---|---|---|---|---| | Audio stays on your device | Yes | No | No | No | | No bot joins your call | Yes | No | No | Yes | | Works fully offline | Yes | No | No | No | | Dual-channel recording | Yes | No | No | No | | On-device AI summaries | Yes | No | No | No | | Native Mac app | Yes | No | No | Yes | Cloud recorders (Otter, Fireflies) upload audio to US servers. This creates legal exposure for everyone on the call, including participants who did not consent to cloud storage. Meeting bots (Otter, Fireflies) are visible to all participants and require permission from the meeting host. Granola is native to Mac and does not use a bot, but it does upload audio for transcription. It does not offer dual-channel recording or on-device transcription. --- ## Use Cases ### Lawyers Client meetings, depositions, and internal strategy sessions involve privileged and confidential information. Cloud recorders create legal risk by uploading audio to third-party servers. Thoth keeps everything on the lawyer's machine, with no third party ever handling the audio. ### Journalists Source interviews often involve sensitive information, off-the-record conversations, and whistleblowers. Uploading these to cloud services violates source protection obligations. Thoth provides a local record with no upload risk. ### Doctors Patient consultations contain protected health information (PHI). Cloud recording tools used without patient consent and BAA agreements create HIPAA liability. Thoth never transmits audio, eliminating cloud compliance concerns for clinical notes. ### Researchers Qualitative research interviews generate transcripts that may include sensitive participant disclosures. Local transcription removes the data governance burden of cloud processing. ### Students Lectures, seminars, and study groups are easier to revisit with an accurate transcript and summary. Recording and transcription run on-device, so notes stay private. --- ## Blog Full blog: https://thoth-app.com/blog/ ### Thoth is coming to iPhone (June 5, 2026) URL: https://thoth-app.com/blog/2026-06-05-thoth-for-iphone/ The private AI scribe for Mac is coming to iPhone and iPad with the same privacy-first architecture: record, transcribe, identify speakers, and summarize on-device, with nothing uploaded. The simplest proof is recording with Wi-Fi off and Airplane mode on. The iPhone app is in open beta on TestFlight while it goes through App Store review. ### Redact names before cloud AI sees them (June 5, 2026) URL: https://thoth-app.com/blog/2026-06-05-on-device-anonymization/ Thoth finds and masks names, places, phone numbers, and emails on-device, then swaps them for tokens before any cloud AI sees the transcript. The summary is de-tokenized locally afterward, so the output reads naturally while the cloud provider never receives the original personal data. ### Thoth for ADHD and autism: what I learned (May 22, 2026) URL: https://thoth-app.com/blog/2026-05-22-thoth-adhd-autism-accessibility/ Thoth was not built with accessibility as the starting point, but a reader review surfaced how much the workflow helps people with ADHD and autism: offloading real-time note-taking, reducing the cognitive load of a meeting, and providing a faithful record to revisit later. ### A band-aid for Parakeet's language drift (May 20, 2026) URL: https://thoth-app.com/blog/2026-05-20-fixing-parakeet-language-drift/ Parakeet TDT v3 drifts into English on spontaneous French speech. A workaround using the top-64 logits applies a soft penalty to English tokens when a non-English language is selected. Not a true fix (the model still lacks language conditioning), but enough to change the recommendation for some non-English audio. ### Parakeet translates French audio (May 19, 2026) URL: https://thoth-app.com/blog/2026-05-19-parakeet-language-drift/ Parakeet TDT v3 was added to Thoth as the fastest transcription model. On clean English audio it nearly matches Whisper Large V3 Turbo while completing 14x faster. On non-English audio, it does something unexpected: it translates rather than transcribes. The model has no language conditioning tokens. When acoustic signal becomes ambiguous (spontaneous speech, regional vocabulary, argot), Parakeet falls back to the language that dominated its training data and outputs English. This is word-for-word involuntary translation. Measured English-language intrusion rates on French recordings: - Children's weather segment (scripted, clear): 0% - Archival 1912 recording (deliberate speech): 0% - Weightlifting documentary 1975: 7.1% - French slang documentary 1981: 18.2% - Picard dialect documentary 1982: 16.7% - Private French-language interview (spontaneous): 31.3% Whisper produced zero English output across all six recordings. Thoth shows a warning before re-transcription with Parakeet when a non-English language is selected. For non-English audio, Whisper Large V3 Turbo is the recommended model. ### Could Thoth come to iPhone? What I found (May 18, 2026) URL: https://thoth-app.com/blog/2026-05-18-thoth-ios-exploration/ Notes on exploring a Thoth port to iPhone and iPad. On-device transcription via WhisperKit is feasible. System audio capture (recording both sides of a call) is not possible on iOS due to platform restrictions. The post covers what would and would not work, and what a limited iOS version might look like. ### Thoth is coming to your menu bar (May 18, 2026) URL: https://thoth-app.com/blog/2026-05-18-menu-bar-mode/ How Thoth's menu bar mode works. The app runs as a compact always-on-top item in the menu bar, enabling one-click recording without switching windows or keeping the main app visible. Designed for users who record frequently and want the smallest possible interruption to their workflow. ### Your meeting audio belongs on your Mac (May 13, 2026) URL: https://thoth-app.com/blog/2026-05-13-why-your-meeting-recorder-shouldnt-upload-your-audio/ Cloud meeting recorders upload audio to US servers. Under GDPR (EU), PIPEDA (Canada), and similar frameworks, recording a conversation and transmitting it to a third-party cloud service may require explicit consent from all participants. In practice, the other people on the call rarely know their audio is being uploaded. Thoth eliminates this exposure by keeping everything local. ### Local vs cloud AI summaries (May 13, 2026) URL: https://thoth-app.com/blog/2026-05-13-local-vs-cloud-ai-summaries/ Detailed comparison of on-device model quality (Qwen 7B) versus cloud AI (Claude Sonnet via BYOK) for meeting summarization. Local models score ~5/10 overall; Claude Sonnet scores ~8.7/10. The gap is largest on decision capture (2/10 vs 8.5/10) and completeness (5/10 vs 9/10). Privacy guarantee is identical for both options: audio never leaves the machine regardless of which AI is used for summaries. ### How I benchmark transcription models (May 10, 2026) URL: https://thoth-app.com/blog/2026-05-10-how-we-benchmark-transcription/ Methodology behind the WER benchmarks published on the site. Three test scripts: clean English, French-accented English, and a French-to-English code-switching script. Ground truth created manually. All models tested on the same 42-minute recording on an M2 MacBook Pro. Published benchmarks (LibriSpeech WER) are consistently 10-30x more optimistic than real-world meeting audio results. --- ## Technical Stack - Language: Swift, SwiftUI - Transcription: WhisperKit (Whisper models via CoreML), FluidAudio (Parakeet TDT v3 and Parakeet EOU via CoreML) - Speaker diarization: Pyannote CoreML - On-device anonymization: local named-entity detection over the transcript - On-device AI summaries: MLX-based local models (Ministral 3B, Gemma 3 12B, and others) - Audio capture: AVAudioEngine, ScreenCaptureKit (system audio on Mac) - Platforms: macOS 26 (Tahoe) and later; iPhone and iPad (open beta) - Distribution: Mac App Store; iOS via TestFlight --- ## FAQ **Does Thoth record both sides of a Zoom or Teams meeting?** Yes, on Mac. Thoth captures your microphone on one channel and the meeting app's system audio on a separate channel, so every participant is recorded without joining a bot to your call. On iPhone and iPad, Thoth records microphone audio only, because iOS does not allow capturing other apps' audio. **Is Thoth on iPhone?** Yes, in open beta on TestFlight: https://testflight.apple.com/join/hu2xjnN7 . The iPhone and iPad app uses the same on-device engine as the Mac app. One Pro purchase unlocks every platform. **Is Thoth free to use?** Thoth is free to try with full transcription, up to 30 minutes per mic recording. Pro unlocks unlimited duration, system audio recording, speaker identification, and unlimited AI enhancements. **Does my audio get sent to the cloud?** No. Transcription runs entirely on your device using WhisperKit and CoreML. Your audio never leaves it. If you use a cloud AI key for summaries, only the text transcript is sent directly from your device to your chosen provider, and on-device anonymization can redact names and other personal data first. **Does Thoth work offline?** Yes. Recording and transcription work with no internet connection. On-device AI summaries also work offline. An internet connection is only needed if you choose to use a cloud AI key. **What Mac models does Thoth support?** Thoth requires macOS Tahoe (macOS 26) or later. Apple Silicon is recommended for on-device AI summaries. Intel Macs are supported for recording and transcription. **What is the difference between Thoth and Granola?** Both are native Mac apps that do not join your meeting as a bot. The key differences: Granola uploads audio to the cloud for transcription; Thoth transcribes entirely on-device. Thoth records both your mic and system audio on separate channels; Granola records mic only and uses your notes as a supplement. Thoth works fully offline; Granola requires internet for transcription. **What is the difference between Thoth and Otter or Fireflies?** Otter and Fireflies both upload your audio to cloud servers and join meetings as visible bots. Thoth does neither. Thoth transcribes on-device with no upload and no bot. Thoth also works offline; Otter and Fireflies require internet for all functionality.