n8n + n8n: Audio & Video Transcription + AI Summary — Whisper + GPT-4o to Notion & Email
You just finished a two-hour discovery call, a 45-minute podcast recording, and a client onboarding video. Somewhere inside those files are the three action items that actually matter — and to find th
You just finished a two-hour discovery call, a 45-minute podcast recording, and a client onboarding video. Somewhere inside those files are the three action items that actually matter — and to find them, someone on your team has to scrub through the audio, pause, rewind, and type notes. That work is expensive, slow, and it never scales. This n8n workflow kills that task entirely: drop any audio or video file into Google Drive, Dropbox, or paste a URL, and minutes later you have a timestamped transcript, an executive summary, key points, notable quotes, and a list of key moments — saved to Notion and sitting in your inbox.
The problem: recordings pile up, insight stays buried
Every technical team generates hours of spoken content weekly — sales calls, standups, user interviews, support escalations, all-hands recordings. The raw files are worthless until a human processes them, and humans are the bottleneck. Manual transcription runs 4x real-time; even a fast note-taker misses the exact quote a stakeholder wanted. Paid transcription SaaS solves half the problem: you get text, but no structured summary, no routing to where your team actually works, and a per-minute bill that grows with volume.
The deeper issue is that transcription and understanding are two different jobs. A transcript is a wall of text. What an ops team or founder needs is the decision that was made, the objection that was raised, the number someone quoted at minute 34. Getting from one to the other is exactly the kind of repetitive, judgment-light reading that a language model does well — if you wire it up correctly. That wiring is what most teams never get around to building.
The solution: a Whisper + GPT-4o pipeline that ends where you work
This workflow chains two AI models inside n8n. OpenAI Whisper handles speech-to-text with word-level timestamps, so every line in the transcript is anchored to a moment in the recording. GPT-4o then reads the full transcript and produces four structured artifacts: an executive summary (2–3 sentences a busy person can read in ten seconds), key points as a bulleted list, important quotes pulled verbatim, and key moments with their timestamps so anyone can jump straight to the relevant part of the source file.
The output doesn't die in a log. It writes to a Notion database — one page per recording, fully searchable and shareable — and simultaneously sends a formatted email to whoever needs it. The trigger is deliberately flexible: a new file in a watched Google Drive folder, a new file in Dropbox, or a direct URL passed to a webhook. You choose the entry point that matches how your team already works, and the rest is identical.
Step-by-step: building it in n8n
The workflow is roughly nine nodes. Here's how they fit together and the configuration that matters at each step.
1. Trigger node. Use the Google Drive Trigger (event: File Created, scoped to a specific folder) or the Dropbox Trigger for cloud storage. For arbitrary URLs, start with a Webhook node (POST) that accepts a JSON body containing the file link. Watching a folder is the lowest-friction option — your team just drops files in and forgets about it.
2. Download the file. If the trigger only hands you metadata, add an HTTP Request node (or the Drive/Dropbox Download operation) set to return the response as a File / binary. Keep the binary property name consistent — data is the n8n default — because the next node reads from it.
3. Transcribe with Whisper. Add the OpenAI node, resource Audio, operation Transcribe. Point its input at the binary property from step 2. In the options, request verbose_json as the response format and enable timestamp granularities at the segment level — this is what gives you the timestamps GPT-4o will reference later. Whisper accepts files up to 25 MB per request; see the pitfalls section for longer recordings.
4. Prepare the prompt. A small Set or Edit Fields node assembles the transcript text (and, if you want, the segment timestamps) into a single variable you'll pass to GPT-4o. This keeps your prompt clean and makes it trivial to tweak later.
5. Summarize with GPT-4o. Add another OpenAI node, resource Chat / Text, model gpt-4o. In the system message, instruct the model to return structured JSON with exactly four keys: executive_summary, key_points (array), important_quotes (array), and key_moments (array of {timestamp, description}). Enable JSON mode / response format json_object so the output parses reliably. Set temperature low (0.2–0.3) — you want faithful extraction, not creativity.
6. Parse the response. Drop in a Code node or use the built-in Structured Output Parser to turn the model's JSON string into real fields you can map. This is the seam where most workflows break, so validate it early.
7. Write to Notion. The Notion node, resource Database Page, operation Create. Map the summary to a title/text property, and push the key points, quotes, and moments into page content blocks. Store the source filename and a link back to the original file so the Notion page is self-contained.
8. Send the email. A Gmail, Send Email (SMTP), or Microsoft Outlook node delivers the same content formatted as HTML. Put the executive summary at the very top — most recipients read only that — with the key moments and Notion link below.
9. Wire the credentials. You need one OpenAI API key (covers both Whisper and GPT-4o), a Notion internal integration token with the target database shared to it, and your email/storage OAuth connections. n8n stores all of these in its encrypted credential store; set them once and every execution reuses them.
Why it's worth building
Speed. A 60-minute recording goes from file to structured summary in a couple of minutes of machine time, with zero human attention. Your team stops trading hours for transcripts.
Consistency. Every recording gets the same treatment — same four artifacts, same format, same place. No more "did anyone write up that call?" Search Notion and it's there.
Cost control. You pay OpenAI's per-minute Whisper rate plus a small GPT-4o completion cost — typically a fraction of per-seat transcription SaaS, and it drops to zero when you're not recording. Because it runs on your n8n instance, there's no third party holding your sensitive call data hostage behind a subscription.
It ends where you work. The whole point is that insight lands in Notion and your inbox automatically — not in yet another dashboard someone has to remember to check.
Common pitfalls (and how to avoid them)
The 25 MB Whisper limit. Long or high-bitrate files will exceed it. Compress audio to a lower bitrate (mono, 64–96 kbps MP3 is plenty for speech) before the transcribe node, or split long files into chunks with an Execute Command/ffmpeg step and merge the transcripts afterward. For most calls under an hour, a quick re-encode is enough.
Video files. Whisper transcribes audio, not video. Extract the audio track first — an ffmpeg step (-vn -acodec libmp3lame) strips the video and shrinks the file dramatically in one move, which also helps with the size limit.
GPT-4o returning malformed JSON. If you skip JSON mode, the model will occasionally wrap its answer in prose and your parser will throw. Always set the response format to json_object and give an explicit schema in the prompt. Add an error branch so a single bad parse doesn't silently drop the whole recording.
Notion property mismatches. The Notion node fails loudly if a mapped property name or type doesn't match your database exactly. Build the database first, then map — and remember to share the database with your integration, or every write returns a permission error.
Token limits on very long transcripts. A multi-hour transcript can approach GPT-4o's context window. If you routinely process long recordings, summarize in segments and then summarize the summaries — a simple map-reduce you can build with a second GPT-4o node over the batched outputs.
Trigger reliability. Polling triggers (Drive/Dropbox) run on an interval, not instantly. If you need real-time processing, prefer the webhook entry point and have your upload tool call it directly.
Ja construimos isso pra voce
Nao comece do zero. O Audio & Video Transcription + AI Summary — Whisper + GPT-4o to Notion & Email e um workflow n8n pronto para instalar — conecta suas ferramentas em minutos, sem codigo.
Instalar por $29.0 →
