Transcript to Summary — Paste, Summarize, Export
Turn an audio recording, video, or a text transcript into a structured AI summary — pick from six purpose-built types (Meeting, Sales Call, Interview, Lecture, Podcast, General). Handles multi-hour recordings with map-reduce architecture, preserves sales-call deadlines verbatim, and never trains on your data. Free 30 minutes at signup, no card.
VexaScribe generates a structured AI summary from any audio, video, or text transcript. Upload an MP3, WAV, MP4, MOV, or any of 17 supported formats up to 5 GB — or paste an existing transcript directly. We transcribe with Whisper Large-v3 (90-95% accuracy on clear audio, 99 languages), then generate a structured summary in one of 6 purpose-built types: General, Meeting, Sales Call, Interview, Lecture, or Podcast. A 60-minute file completes in 5-10 minutes. Under the hood: map-reduce architecture with semantic chunking — embeddings detect natural topic shifts in long recordings, each section is summarized in parallel, then consolidated. Multi-hour recordings preserve their structure instead of getting flattened to a paragraph. Summaries can be generated in any of 99 target languages regardless of source language. Free tier is 30 minutes at signup — no credit card, all 6 summary types included, no training on your data ever.
Not quite what you're looking for?
- Want an AI note-taker that joins meetings automatically? → /ai-note-taker
- Want a dedicated AI meeting minutes generator with action items and decisions? → /ai-meeting-minutes-generator
- Just want a summary of a YouTube video? → /video-summarizer
Why summarize transcripts at all?
Raw transcripts are searchable but unreadable. Summarization is what turns 90 minutes of audio into a 90-second decision — and the time savings are measurable.
Numbers verify what every knowledge worker already feels: meetings and unfiltered transcripts are a productivity tax. AI summaries are the recovery mechanism. See how we verify these stats.
What is an audio summarizer — and how is it different from a text summarizer?
A text summarizer (Quillbot, ChatGPT, etc.) takes text you paste in and condenses it. An audio summarizer starts a step earlier: it transcribes the audio first, then summarizes. That upstream step matters because audio has no punctuation, no paragraph breaks, and no natural formatting — the model that writes the summary has to infer sentence boundaries, speaker turns, and topic shifts from raw word sequences. A generic text summarizer pasted with an unformatted transcript will average those cues away; a purpose-built audio summarizer is trained and prompted around them.
VexaScribe handles both steps in one pipeline: Whisper Large-v3 for the transcription (with word-level timestamps and speaker diarization), then a map-reduce summarization pass that uses semantic chunking to find natural topic boundaries before summarizing each segment. The final output is not a condensed paragraph — it is a typed, structured document matched to your recording format: meeting notes, show notes, study guide, or sales call report.
If you already have a transcript, you can paste it directly into the summary tool and skip the transcription step — but the pipeline is designed to work best end-to-end so speaker labels carry through to the summary.
Podcast summarizer
The Podcast summary type is built for host + guest formats. It extracts speaker profiles (name, role, what they argued), per-speaker discussion points with the evidence they used, memorable key insights suitable for social posts or show notes, explicit agreements and disagreements between speakers, and all book/tool/resource recommendations with who made them and why.
Output is formatted as ready-to-publish show notes — executive summary at the top, chapters with timestamps, guest highlights, recommendations list at the bottom. A 60-minute episode typically summarizes in under 2 minutes. See the sample output for a worked example.
Lecture summarizer
The Lecture summary type is designed for university lectures, training sessions, and tutorials. It outputs key concepts with 2–3 sentence explanations (the ideas students need to deeply understand), every analogy and case study with what it illustrates, a terminology glossary (jargon and acronyms, explicitly separated from key concepts), 3–7 main learning takeaways, review questions with answer hints, and further-reading suggestions the lecturer mentioned.
This is the format students use for exam prep and the format instructors use to generate supplemental materials. The distinction between key concepts (deep understanding required) and terminology (definitions) is enforced by the prompt — so the glossary does not get bloated with bullet points about ideas covered in the concepts section.
How VexaScribe AI summary works
Most AI summarizers feed the entire transcript into one call and accept whatever comes back. That works on a 5-minute voice memo and falls apart on a 90-minute meeting. VexaScribe uses a map-reduce pipeline with semantic chunking — so even multi-hour recordings preserve their structure instead of getting flattened to a paragraph.
- 1
Semantic chunking
For long transcripts, we use OpenAI embeddings to find topic boundaries in the conversation — measuring sentence-to-sentence similarity, then locating 'valleys' in the similarity curve that mark topic shifts. Chunks get cut at those natural boundaries, not at arbitrary length thresholds.
- 2
Map phase
Each chunk goes to a current-generation OpenAI model in parallel, extracting chapters and type-specific structured data. Overlap segments from the previous chunk are included so context isn't lost across boundaries.
- 3
Reduce phase
A final call consolidates all chunk results into one coherent summary — deduplicating action items, merging only chapters that are genuinely the same topic, and producing the final structured output for your chosen summary type.
Why this matters in practice: a 2-hour planning meeting summarized naively flattens 4 distinct decisions into one vague paragraph. With semantic chunking, the AI sees four topic segments, summarizes each separately, then consolidates — preserving the actual structure of the conversation. Powered by current-generation OpenAI models for both map and reduce phases.
Pick your summary type (6 options)
The right summary depends on what kind of recording you uploaded. VexaScribe ships six purpose-built templates — each returns a different structured field set. Pick one before you click Generate, and switch any time without re-uploading.
General
Default for any audio without specific context — voice memos, briefings, mixed content
Sections included
- Topics — 3-10 high-level subjects discussed
Use when the recording doesn't fit a specialized category — flexible high-level structure.
Meeting
Team meetings, standups, planning sessions, retros
Sections included
- Action items — task + assignee + deadline. Detects both volunteers ('I'll handle that') AND assignments ('Can you do X?')
- Decisions — 3-10 distinct decisions per meeting (what was chosen — separated from action items)
- Open questions — only questions that stayed unresolved by the end (filters out questions that got answered)
- Blockers — risks, dependencies, impediments raised during the meeting
Strict separation of decisions vs action items: a decision is what was chosen, an action item is who does what.
Sales Call
Sales discovery, demo, negotiation, closing calls
Sections included
- Action items — with strict rule: deadlines preserved in the speaker's exact words (never converted to a calendar date)
- Client needs — need + priority (high/med/low) + specific evidence quote in the client's own words
- Objections — pushback raised, seller response, resolution status (yes/partially/no)
- Competitor mentions — competitor named + what was said + positioning (positive/negative/neutral)
- Pricing discussions — every dollar amount, plan, discount discussed + specific outcome
- Deal next steps — owner + deadline + priority — pipeline-style
- Sentiment — one-liner read with reasoning ('Positive — client agreed to POC and expressed urgency')
- Deal stage — Discovery / Demo / Negotiation / Closing / Closed Won / Closed Lost
- Qualification signals — BANT-style — Budget, Authority, Need, Timing observations
Most field-rich of the six. The deadline-preservation guardrail (deadlines kept in the speaker's exact words, never converted to fabricated calendar dates) is unique to this type — verified in our internal testing.
Interview
Candidate interviews, panel discussions, hiring loops
Sections included
- Notable exchanges — pivotal Q&A moments — topic, summary, speakers, timestamp
- Strengths — positive observations about the interviewee
- Concerns — red flags, gaps, weaknesses noted
- Overall assessment — one-paragraph synthesized evaluation
Designed for hiring loops — surfaces strengths, concerns, and a synthesized assessment.
Lecture
University lectures, training sessions, tutorials
Sections included
- Key concepts — big ideas, theories, frameworks — with 2-3 sentence explanations (the things students need to deeply understand)
- Examples given — every analogy or case study + what it illustrates + the specific teaching insight it conveys
- Terminology — vocabulary, jargon, acronyms — glossary style (explicitly separated from key concepts)
- Takeaways — 3-7 main learning outcomes
- Review questions — 3-7 study questions, each with an answer hint
- Further reading — papers, books, topics the lecturer pointed to for deeper exploration
Built for studying — review questions, terminology glossary, and further-reading suggestions all included.
Podcast
Conversational shows, host + guest formats
Sections included
- Speaker profiles — name + role per speaker (extracted from intros)
- Discussion points — per speaker, per topic, with the specific evidence/numbers the speaker used
- Key insights — memorable stats, claims, predictions, surprising data (the 'tweetable' moments)
- Agreements and disagreements — explicit moments where speakers aligned or diverged + which speakers
- Recommendations — books, tools, resources mentioned + who recommended + why
- Guest highlights — 2-4 standout moments per guest (specific, not generic)
Tailored for show-notes publishing — captures per-speaker positions, recommendations, and standout moments.
Quality is consistent across types — instruction-tuned models match human-written quality on standardized news benchmarks (Zhang et al., TACL 2023). The differentiator is which structured fields surface what you actually need from the conversation. Picking a tool? See our honest comparison of 10 podcast transcription tools (Descript, Castmagic, Otter, Rev, and more).
See it on real recordings
One short before/after example for each summary type — Meeting, Sales Call, Lecture, Interview, Podcast, and General. Each side capped at ~80 words so you can scan all six in two minutes.
Audio transcript excerpt
…so the renewal pricing came back at 18% up year-over-year, which Priya flagged as risky given Acme's Q2 results. Tom suggested we counter at 9% and lock for 24 months. Priya owns the redline by Friday. We also agreed to pause the SSO migration until legal signs off — Rahul will confirm the lawyer's bandwidth Monday. No decision on the Slack-vs-Teams switch; bumped to next week's all-hands…
AI summary output
Audio transcript excerpt
…yeah, the pain is mostly that we run a Monday revenue meeting and Sarah pulls the numbers from three different dashboards. It takes her like three or four hours every week. Budget-wise we have something allocated for next quarter but Q2 is locked. Decision-wise, I'm the one signing but our CFO needs to bless anything over $50K annual. Send me an ROI model with realistic numbers and I'll loop her in next week…
AI summary output
Audio transcript excerpt
…so when we talk about externalities in welfare economics we mean a cost or benefit imposed on a third party that isn't reflected in the market price. Pollution is the classic negative case. Pigou's solution in 1920 was a tax equal to the marginal external cost — the so-called Pigouvian tax. Coase pushed back in 1960: if transaction costs are low, parties can bargain to an efficient outcome regardless of who holds the property right…
AI summary output
Audio transcript excerpt
…my approach to a hard reorg is to compress the information gap. The week before is the hardest part — when leadership knows and the team doesn't. We had to lay off forty people last year. I made the decision Monday, communicated Tuesday, supported them with severance and intros Wednesday. The trust we kept came from speed, not from the message…
AI summary output
Audio transcript excerpt
Host: …and what most founders miss is that the first 50 customers aren't a market — they're a focus group. Guest: I'd push back slightly. They're a market if you've already nailed the ICP. The trap is when you confuse the loudest five for the average. Host: Fair. So how do you avoid that? Guest: Write the ICP in one sentence and tape it to your monitor…
AI summary output
Audio transcript excerpt
…just dumping some thoughts about the offsite. The hike was great, dinner was fine but service was slow. The most useful session was the founder Q&A — three things came up worth following up on: the pricing experiment timeline, the hire for ML platform, and whether we want to go back to the same venue next year…
AI summary output
Beyond summaries: ask follow-up questions with AI Chat
A summary gives you the gist. AI Chat lets you ask anything specific the summary missed — a natural-language Q&A interface scoped to one transcript, with citation chips that link back to the exact second in the audio where each quote came from.
How it complements summaries
- ● The summary gave you the key decisions — chat lets you ask “why did we decide that?” and get the reasoning quoted from the discussion
- ● The summary listed action items — chat lets you drill into “did anyone push back on the deadline?”
- ● The summary skipped something — chat lets you find it: “did anyone mention Q4 revenue numbers?”
- ● The summary is in English — chat lets you ask in your native language and get answers in it, with original-language quotes preserved
What's structurally different
- ● Citations are validated against the source. Quotes the AI proposes that don't actually appear in the transcript are dropped before reaching you.
- ● Timestamps are real. Click a citation chip — the audio player seeks to that exact moment for verification.
- ● Constrained to the recording. If something isn't in the audio, the AI says so rather than inventing.
- ● Multi-hour recordings handled via semantic retrieval. Tested on 6+ hour recordings with 21 speakers.
Action items extraction — what gets captured
The Meeting and Sales Call summary types extract action items as task + assignee + deadline. The prompt explicitly looks for two linguistic patterns: volunteer signals (“I'll handle that,” “I can take this on”) and assignment signals (“Can you do X?”, “John, can you handle this?”). Hedged language is filtered out by design — “maybe we should...” doesn't get flagged, to avoid false positives.
Transcript excerpt (Meeting)
Tom: So we need someone to own the redline by Friday. Priya: I can take that — I'll have it back to legal by Thursday afternoon. Rahul: Should I confirm the lawyer's availability? Tom: Yes, ping them Monday and let us know. Sarah: Let's circle back on the Slack-vs-Teams decision next week — I don't think we have time today.
Extracted action items
| Task | Assignee | Deadline | Detected as |
|---|---|---|---|
| Send redline to legal | Priya | Thursday afternoon | Volunteer pattern ('I can take that') |
| Confirm lawyer's availability | Rahul | Monday | Assignment pattern ('Yes, ping them Monday') |
How long should a transcript summary be?
Industry convention: a summary is typically 5-15% of the source transcript length. Short summary ~2-3%, executive summary ~5%, detailed summary ~10-15%. The right answer depends on who's reading it and what they need to do next.
VexaScribe doesn't force a target length — the summary is structured around fields (executive summary, chapters, action items, decisions, etc.) and the field content scales naturally with the source. The table below gives you a planning reference for the most common recording lengths.
| Source length | Short summary | Executive | Detailed | When to use which |
|---|---|---|---|---|
| 60-min meeting (~9,000 words transcribed) | 180-270 words (2-3%) | 450 words (5%) | 900-1,350 words (10-15%) | Short for Slack post-meeting; Executive for stakeholder digest; Detailed for project documentation |
| 30-min sales call (~4,500 words) | 90-135 words | 225 words | 450-675 words | Short for CRM Activity note; Executive for deal review; Detailed for handoff to AE / CSM |
| 60-min podcast (~9,500 words) | 190-285 words | 475 words | 950-1,425 words | Short for social-media teaser; Executive for show notes header; Detailed for full SEO-friendly show notes |
| 90-min lecture (~13,500 words) | 270-405 words | 675 words | 1,350-2,025 words | Short for quick review; Executive for study guide; Detailed for exam prep |
Accuracy and hallucination — how we handle it
Any generative AI summary can hallucinate — write things that weren't in the source. Our mitigations are at the prompt level and the architecture level. Below: the designed safeguards, what we've actually tested, and the honest limits.
What we've verified in internal testing
Small-sample testing (2 real transcripts, 4 summary runs). Two real data points beat a single “X% accurate” vendor claim. Verification date: 2026-06-29.
- Action item extraction recall: On a 10-minute board meeting with 7 hand-annotated ground-truth action items, the Meeting summary extracted 5 of 7 (71% recall) with 0 hallucinated items. Misses were borderline cases (one was arguably-not-an-action-item). Zero false positives matters more than total recall for trust.
- Sales-call deadline preservation: On a transcript with relative deadlines (“tonight tomorrow night,” “before tuesday january 27th”), the sales_call summary returned 0 fabricated calendar dates — deadlines stayed as “before January 27th” (verbatim). The same content summarized as Meeting DID fabricate dates (became “2026-01-22” and “2026-01-27”). This is a real, measurable difference between summary types.
- Action item assignee accuracy: On the 10-minute board meeting, 4 of 5 assignees were correctly named. One failure: an action item was assigned to “Shelley Scott” when the actual speaker was “Neil” — diarization mislabeled the speaker, and the summary inherited that label.
Anti-fabrication instructions in every prompt
All six summary types are explicitly instructed to never fabricate content not present in the transcript. The map and reduce phases both receive this guardrail. In our testing, zero action items were hallucinated.
Sales-call deadline preservation (unique to this type)
Sales Call summaries are explicitly instructed to preserve deadlines in the speaker's exact words. If a prospect says “next Thursday,” the summary says “next Thursday” — we don't convert relative dates to specific calendar dates. The instruction is repeated three times across the map prompt, reduce prompt, and system role to maximize compliance.
Honest limitation: this guardrail currently applies to sales_call summaries only. Meeting summaries may still convert relative dates to calendar dates — if you're summarizing a meeting where literal deadline wording matters, pick the Sales Call summary type even if the recording is internal. The meeting-prompt guardrail is on our backlog.
Map-reduce architecture reduces drift
Because each chunk is summarized in isolation against its own source segment (with overlap context), the model can't fabricate by “blending” far-apart parts of the transcript. The reduce phase consolidates from already-grounded chunk summaries rather than re-summarizing from scratch.
- ● No automatic entity deduplication pass. On long recordings, if a speaker is referred to by different names — “Dr. Smith” early on and “Mike” later — they may appear as two separate people in the summary's speaker fields. The model is sometimes good at unifying contextually within a single chunk, but cross-chunk unification is unreliable. We're improving this; it's not in production yet.
- ● No automatic verbatim cross-check pass against the transcript. We rely on prompt-level anti-fabrication instructions; we don't run a second AI pass to verify every quote.
- ● Assignee accuracy depends on diarization quality. Well-mic'd recordings with distinct voices produce clean assignee fields. Single-microphone conference rooms can blur speakers together, in which case assignees may inherit incorrect diarization labels (“SPEAKER_00”) or be mislabeled.
- ● Meeting-summary deadlines may be converted to calendar dates. See above — this guardrail exists only in sales_call mode today.
- ● High-stakes content (legal, medical, journalism quote-checking) still warrants human review. AI summaries are designed for everyday meetings, podcasts, lectures, and sales calls — the volume problem.
Why not just paste the transcript into ChatGPT?
Honest answer: ChatGPT works for one-off summaries when you already have the transcript text. For audio-in, repeatable, structured workflows it's the wrong tool — but it's a real option for occasional needs. The honest scorecard:
| Criterion | ChatGPT (paste a transcript) | VexaScribe | Advantage |
|---|---|---|---|
| Audio / video input | No — text only | Yes — 17 formats up to 5 GB | Us |
| Long-recording handling | Context-window-limited on multi-hour transcripts | Semantic chunking + map-reduce — multi-hour preserved | Us |
| Structured output | Requires custom prompt every time | 6 pre-built schemas (Meeting / Sales / Interview / Lecture / Podcast / General) | Us |
| Sales-deadline literal preservation | No guardrail | Explicit instruction — 'next Thursday' stays 'next Thursday' | Us |
| Multilingual summary (transcribe X, summarize Y) | Yes | Yes — 99 input languages × 99 output | Parity |
| Export to Markdown / DOCX / Notion / Slack | Manual | 1-click | Us |
| Training on your data | ChatGPT Plus trains on chats unless opted out | Never — explicit policy | Us |
| One-off transcript you already have | Fine — paste and prompt | Overkill — sign-up required | ChatGPT |
The honest framing: ChatGPT is a great general-purpose tool for ad-hoc transcript summaries when you already have the text and don't need a specific schema. VexaScribe is built for the audio-in, content-typed, repeatable case — and is the right tool when you want to skip the “design a prompt every time” cost.
AI summary vs human-written: when each wins
AI wins on cost, speed, and scale. Humans still win on legal liability, deep cultural nuance, and stakes-where-a-mistake-is-fatal. The honest scorecard:
| Criterion | AI summary | Human-written | Winner |
|---|---|---|---|
| Cost per 60-min summary | ~$0.00–$0.30 | $30–$80 (freelance) | AI |
| Turnaround time | ~15 seconds | 2–24 hours | AI |
| Accuracy on standard transcripts | On par with human-written quality on standardized news benchmarks (Zhang et al., TACL 2023). Designed safeguards: explicit anti-fabrication instructions; sales-call deadlines preserved in speaker's exact words. | ~96–98% with domain-expert reviewer | Human (narrowly) |
| Cultural / idiomatic nuance | Misses sarcasm, regional idioms | Strong when reviewer shares context | Human |
| Legal / medical liability | Not certified; output is not auditable testimony | Trained transcriptionist + signed attestation | Human |
| Scale (1,000+ transcripts/week) | Trivial | Requires a team | AI |
| Format consistency across runs | High — deterministic templates | Variable across humans | AI |
| Long-tail languages (e.g., Welsh, Swahili) | Strong on top 25, weaker on long tail | Depends on reviewer availability | Tie / depends |
AI wins for everyday meetings, podcasts, and lectures — the volume problem. Humans still win when a single misquote can land you in court or harm a patient: legal depositions, medical records, high-stakes journalism. VexaScribe's stance: ship AI summaries with explicit anti-fabrication instructions and sales-call deadline preservation, route the 5% of high-stakes cases to human review. See our editorial review process.
Summarize in any language — or translate while you summarize
VexaScribe transcribes in 99 input languages and can deliver the summary in any of 99 target languages — input and output are independent. Upload a Spanish meeting recording and get English action items, or vice versa, in a single workflow.
Common pairings:
- • Spanish meeting recording → English action items (US team distribution)
- • German lecture recording → English chapters + key concepts (international students)
- • Japanese podcast → English show notes (cross-market publishing)
- • Portuguese interview → English quotes + themes (research synthesis)
For pure audio translation without summarization, see transcribe and translate audio.
Export the summary anywhere
The summary downloads as Markdown, DOCX, or plain text — and copy-to-clipboard preserves Markdown formatting so you can paste cleanly into Notion, Obsidian, Google Docs, or Slack. The full transcript exports separately as TXT, DOCX, SRT, VTT, or JSON.
| Integration | Supported export | Setup |
|---|---|---|
| Markdown file | .md with frontmatter + headings | Direct download, no setup |
| DOCX (Word) | Word-compatible with styles | Direct download, no setup |
| Plain text | .txt for any text editor | Direct download, no setup |
| Notion (paste) | Copy to clipboard with Markdown formatting | Paste into any Notion page |
| Obsidian (paste) | Markdown with wiki-link compatible headings | Paste into vault |
| Google Drive (paste) | Copy to clipboard, paste into Doc | Manual paste |
What happens to your recording
Audio uploads are encrypted in transit (TLS 1.2+) and at rest. We don't train AI models on your recordings, and you can delete files any time.
- TLS 1.2+ in transit, encrypted at rest in AWS eu-west-2.
- No model training on customer data — your recordings, transcripts, and summaries are yours.
- Self-serve deletion any time from your dashboard.
- Account deletion purges all recordings, transcripts, and summaries.
Full details in our privacy policy.
How to generate a transcript summary
Three steps from audio file to structured AI summary.
- 1
Upload audio or video
Drag-drop an MP3, WAV, M4A, MP4, MOV, or any of 17 supported formats up to 5 GB. Source language is auto-detected across 99 supported languages.
- 2
Pick a summary type
Choose from 6 types — General, Meeting, Sales Call, Interview, Lecture, or Podcast. Each generates a different structural template. Switch types and regenerate without re-uploading.
- 3
Edit and export
Review the transcript and summary side-by-side in the synced editor. Correct anything, then download the summary as Markdown or DOCX, or copy to clipboard for Notion / Drive / Slack.
Transcript Summary — Frequently Asked Questions
How does transcript summary work on VexaScribe?
Upload an audio or video recording (MP3, WAV, M4A, MP4, MOV, and 12 other formats up to 5 GB). VexaScribe transcribes the audio with Whisper Large-v3, then generates an AI summary tailored to your chosen type — General, Meeting, Sales Call, Interview, Lecture, or Podcast. A 60-minute file typically completes in 5-10 minutes including both transcription and summary.
What audio and video formats are supported?
MP3, WAV, M4A, FLAC, OGG, AAC, AIFF, WMA, AMR, OPUS for audio, and MP4, MOV, AVI, MKV, WebM, FLV, WMV for video. Files can be up to 5 GB and 10 hours long. For video files, audio is extracted automatically.
What summary types are available?
Six purpose-built types — General, Meeting, Sales Call, Interview, Lecture, and Podcast. Every summary includes an executive summary, chapters, and key quotes. Each type then adds specialized sections: Meeting adds action items (task, assignee, deadline), decisions (filtered to exclude action items), unresolved open questions, and blockers. Sales Call adds action items with deadlines preserved in the speaker's exact words, client needs with priority and evidence quotes, objections with resolution status, competitor mentions with positioning, pricing discussions with outcomes, deal next steps, sentiment with reasoning, deal stage (Discovery / Demo / Negotiation / Closing / Closed Won/Lost), and BANT qualification signals. Lecture adds key concepts, examples with teaching insight, terminology glossary, takeaways, review questions with answer hints, and further reading. Interview adds notable exchanges, strengths, concerns, and an overall hire assessment. Podcast adds speaker profiles, discussion points per speaker, key insights, agreements and disagreements, recommendations, and guest highlights. General adds a topics list.
How does VexaScribe handle long recordings without losing structure?
Map-reduce architecture with semantic chunking. For long transcripts, we use OpenAI embeddings (text-embedding-3-small) to detect natural topic boundaries — sentence-to-sentence similarity 'valleys' that mark topic shifts. Each chunk is summarized in parallel with overlap context from the previous chunk, then a final reduce phase consolidates everything into one coherent summary, deduplicating action items and merging only chapters that are genuinely the same topic. Multi-hour recordings preserve their structure instead of getting flattened to a single paragraph.
How accurate are AI summaries, and can they hallucinate facts?
On standardized news benchmarks, instruction-tuned LLM summaries are judged on par with human-written ones (Zhang et al., TACL 2023). In our own internal testing on a 10-minute board meeting with 7 hand-annotated action items, the Meeting summary extracted 5 of 7 (71% recall) with zero hallucinated items. On the same transcript, sales_call mode preserved relative deadlines verbatim ('before January 27th') while Meeting mode fabricated calendar dates from the same input — a real, measurable difference between summary types. Safeguards: all six types are explicitly instructed not to fabricate content not in the transcript, and sales_call mode has a triple-repeated deadline-preservation guardrail. Honest limitation: the deadline guardrail exists only in sales_call mode — meeting summaries may still convert relative dates. For legally-sensitive content, humans still win — see the AI vs human comparison section.
Is my recording private — do you train on it?
No. Audio files transit over TLS 1.2+ encryption and are stored encrypted at rest in AWS eu-west-2. We don't use customer data to train AI models. Self-serve deletion any time from your dashboard. Account deletion purges all recordings, transcripts, and summaries.
Which languages does the summary support?
VexaScribe transcribes in 99 languages with automatic language detection. Summaries can be generated in any of 99 target languages — feed a Spanish meeting recording and get English action items, transcribe Turkish and summarize in English, etc. Input language and output language are independent.
What does the free tier include?
30 minutes of transcription on the free preview, with summary generation included. Paid plans: Starter $2/mo (200 min), Basic $5/mo (1,000 min), Pro $10/mo (2,500 min), Studio $20/mo (6,000 min). All plans include all 6 summary types and all export formats.
Why not just paste the transcript into ChatGPT?
ChatGPT works for one-off summaries when you already have the transcript text. VexaScribe is the integrated workflow for repeatable use: upload audio once and get a transcript plus six pre-built summary schemas; no need to design a custom prompt every time. We chunk long recordings semantically (ChatGPT is context-window-limited on multi-hour audio), preserve sales-call deadlines in the speaker's literal words, and never train on your data. ChatGPT Plus trains on chats by default unless you opt out; we don't, ever.
What does the AI summary not handle perfectly?
Two honest limits worth knowing. First, entity unification across long recordings — if a speaker is referred to as both 'Dr. Smith' and 'Mike' at different points, they may appear as two separate people in the summary's speaker list. We don't run a name-canonicalization pass yet. Second, action item assignees depend on diarization quality — well-mic'd meetings with distinct voices produce clean assignee fields, but single-microphone conference rooms can blur speakers together and assignees may fall back to whoever is explicitly named in the spoken sentence. Both are areas of active improvement.
Which summary type should I pick for a 60-minute team meeting versus a 90-minute lecture?
Team meeting: pick Meeting — surfaces action items with assignees, decisions, and blockers in a structured layout. Lecture: pick Lecture — generates chapters with timestamps plus key concepts, terminology glossary, and review questions for study. Podcast: pick Podcast for publishable show notes with speaker profiles. Sales call: pick Sales Call for objections, deal stage, BANT signals, and deadlines preserved literally.
Related tools
Video summarizer
Start from a video (or YouTube URL) instead of a transcript
Video to notes
Structured working notes (5 formats) instead of a short summary
Transcribe audio
Get a transcript first, then summarize
Whisper transcription
Whisper engine that powers transcripts
AI Meeting Minutes Generator
For formal minutes (board, project, sales, non-profit) with template structure — different from informal summaries
MP3 to text
Convert podcast and voice memo MP3s
Podcast transcription
Show notes, episode transcripts, SRT
Lecture transcription
AI study guides from lecture recordings
Sermon transcription
Sermon-to-blog pipeline — summary is half the workflow
Transcription for qualitative research
Methodology + IRB + CAQDAS for researchers
Transcribe & translate
99 source × 99 target languages
SRT generator
Subtitle files for video summaries
Best podcast transcription tools
10-tool honest comparison for podcasters
Sales call transcription
10-tool ranked comparison — Gong, Chorus, Fireflies, and where VexaScribe fits
Voicemail transcription
iPhone, Android, carrier and file-based options compared
Meeting transcription
The upstream workflow — the Meeting summary template starts here