How do I search hours of my own recordings by what was said and who said it?

To search hours of your own recordings (interviews, calls, lectures, meetings) by what was said and who said it, the recordings have to live in one place that keeps three things for every passage: the words, the speaker and the timestamp. On Audioscrape you send the recordings to your private workspace, from the website, from ChatGPT or Claude, or through the API; each one is transcribed with the speakers separated; you name the people who recur once; and from then on you can search the whole archive by exact words, by topic, or by speaker, and every result links to the second it was said.

Why a folder of transcripts is not enough

A transcription tool turns one file into one text. That works for a single recording. With fifty, the question changes from “what was said in this file” to “where did anyone say X, and who said it”, and a folder of text files cannot answer that: the search is per file, Speaker 1 in Monday’s call is unrelated to Speaker 1 in Tuesday’s, and a quote without a timestamp cannot be checked against the audio.

How to do it with Audioscrape

1. Send the recordings. Any of these ends up in the same workspace:

• On the website, drag files onto the upload page.
• In ChatGPT, with the Audioscrape plugin installed, attach the recording to the chat and ask to transcribe it.
• In Claude, share the file from Google Drive, Dropbox or OneDrive with “anyone with the link” and paste the link; Claude cannot pass attachments to a connector.
• From a script, POST /api/datasets/{id}/items with the file or a URL (Upload API).

Audio and video both work: mp3, wav, m4a, ogg, flac, aac, opus, mp4, mov, webm and mkv. A file already in the workspace is recognized and not transcribed or counted twice.

2. Let it transcribe. Each recording is transcribed in the background with speakers separated and every segment timestamped. Most finish within minutes; long recordings take longer.

3. Name the voices once. Open a finished transcript and name the people who will come up again: you, your team, a regular client, a recurring guest. The workspace keeps those identities, so later recordings come back with the names attached wherever the voice match is confident. Details: identifying the same speakers across recordings.

4. Search. Use exact words for a phrase or a name, and semantic search for a topic said in different words. Filter by speaker to see only what one person said: speaker: "Maria Chen" pricing. Each hit shows the speaker, the recording, the timestamp and a link that opens playback at that moment.

5. Or ask. In ChatGPT or Claude, ask in plain language: “What did Maria say about pricing across all our calls since September?” The assistant searches your workspace and answers with the passages, the speakers and the links. See searching your recordings inside ChatGPT or Claude.

What you get

• One searchable archive across every recording, not one transcript per file.
• Speaker names that carry over to new recordings once you have named a voice.
• A timestamp and a link to the audio on every result, so any quote can be checked in seconds.
• Your recordings searchable alongside public podcasts, council meetings and court hearings when you want both in one answer, and on their own when you do not.
• Privacy: recordings stay in your workspace, visible only to its members and the assistants or keys you connect, and are never used to train models without your permission.

Limits and honest caveats

• Audio quality decides accuracy. Distant microphones, phone lines and crosstalk lower both the transcript and the speaker separation.
• Speaker names are machine labels. The first recording of anyone is anonymous until you name them, and a similar voice on a poor line can be mislabeled; correct it in the transcript and the fix carries forward.
• Plans: Free includes 5 hours of transcription a month with recordings up to 3 hours each. Starter ($35 a month) includes 20 hours and Pro ($129 a month) 100 hours, with no per-file limit. See pricing.
• Recording people can require their consent, and voice identification is biometric data in some places. Upload only audio you have the right to process.

When something else fits better

If you need one file turned into text once, any transcription tool, or a model running on your own computer, will do and may cost less. If you need a bot that joins live calls and writes notes, that is a meeting assistant, which Audioscrape is not. Audioscrape is for the moment the recordings pile up and the question becomes where something was said, and by whom. A side-by-side with a meeting assistant: Audioscrape vs Otter for an archive.