Guides · Transcripts
Use Whisper transcripts with Daytapes (.srt, .vtt, .json)
Make a transcript with OpenAI Whisper, whisper.cpp or a local app, then attach it to your audio in Daytapes. Formats, limits and when local makes sense.
Whisper is a family of speech-to-text models from OpenAI. You can run it on your own computer and get a transcript file, usually .srt, .vtt, .txt or .json. Attach that file with your audio in Daytapes, and the page gets a transcript that plays along with the recording.
You never have to do this. Daytapes transcribes audio by itself. A Whisper file is for people who already use Whisper, who want audio to stay on their own machine for the first pass, or who have a transcript they like. This is a research-based guide, built from the projects' own pages. We haven't run speed or accuracy tests.
What "Whisper" can mean
People use one word for several things. Know which one you have.
- OpenAI Whisper (open source). The original Python project. Its GitHub page says the code and model weights use the MIT license. It lists six model sizes, from tiny (39M parameters) to large (1550M), plus a turbo model (809M). It needs ffmpeg, Python and PyTorch.
- whisper.cpp. A plain C and C++ port. Its README says it has no dependencies, runs on CPU, and treats Apple Silicon as "a first-class citizen," with Metal and Core ML support. It's also MIT-licensed.
- Local apps. Many desktop and phone apps wrap Whisper or whisper.cpp in a simple window. Features and prices vary. Check the app's own page.
- OpenAI's hosted API. A paid cloud service. OpenAI's speech-to-text guide, read November 14, 2026, recommends a model named
gpt-transcribe, and sayswhisper-1, along with three other earlier models, is deprecated and "shuts down February 26, 2027." The guide says files can be up to 25 MB.
If you plan to build on the hosted API, check the model name before you start. A date like that catches people out.
Which file to make
Daytapes accepts transcript files in .txt, .srt, .vtt and .json, attached with the audio. Here is how they compare.
| Format | Has timing | Best for |
|---|---|---|
| .srt | Yes, a start and end per line | The safe default. Widely supported subtitle format. |
| .vtt | Yes | Same idea as SRT. WebVTT is a W3C text format for captions, and files begin with the word WEBVTT. |
| .json | Often, with more detail | Keeping words and timing together. Contents depend on the tool. |
| .txt | No | Quick reading. Least precise for syncing to audio. |
The Whisper command line, per its source code, offers --output_format choices of txt, vtt, srt, tsv, json, jsonl and all, and the default is all. That means a plain run writes every format into a folder. Pick the .srt or .vtt for Daytapes, and keep the rest.
A basic run
If you install the open-source package, a command like this writes transcripts next to your audio:
whisper "2026-11-14 standup.m4a" --model turbo --output_format srt --output_dir transcripts
Check whisper --help for the options in your installed version. Our sources didn't list every flag, and tools change.
For whisper.cpp, the README we read didn't spell out its subtitle options. Run its help command to see what your build offers.
Does Whisper label speakers?
Mostly no. The open-source Whisper command line gives you words and timing, not "who said it." whisper.cpp lists "speaker-turn markers (tinydiarize)" as an experimental option. OpenAI's API guide mentions a diarized_json response format that returns segments with speaker, start and end, but it also lists the related diarize model as deprecated.
Don't worry about it. Daytapes works out who's talking and puts names on the page. You only need to check the names once the page appears. If a person is mislabeled, correct the name on the page. If you want to know what the word means, see "diarization" in the glossary.
When local transcription makes sense
Local isn't better in every case. It's a trade.
Good reasons to transcribe locally
- The audio shouldn't leave your machine yet. A local run keeps the first pass on your own computer. You can read it, trim it and decide what to share. Note that attaching it to Daytapes still sends the audio there.
- You already have a workflow. If your files already go through Whisper, attaching the output costs nothing.
- You want to pick the model. Larger models are slower and use more memory. whisper.cpp's README lists roughly 75 MiB of disk for tiny up to about 2.9 GiB for large.
- You're offline. A laptop on a plane can still transcribe.
Good reasons not to
- Setup takes time. Python, ffmpeg and PyTorch, or a compiled program, are real work for non-technical people.
- Speed. Big models are slow on older hardware.
- Accuracy isn't guaranteed. Whisper can mishear names, numbers and jargon, and can drift on noisy audio. Spot-check before trusting a number.
- Labels and polish. You'll still need speaker names and a summary, which Daytapes adds.
What the hosted API keeps
If you use OpenAI's hosted API instead of local Whisper, read its data rules. OpenAI's "Your data" page, read November 14, 2026, says abuse monitoring logs are "retained for up to 30 days" by default, unless longer retention is required by law. It also says data sent to the API is not used to train its models by default. Approved customers can ask for zero data retention. Policies change, so read the page before you send sensitive audio.
Put it together: audio plus transcript in Daytapes
- Name the files alike. Match the base name, for example
2026-11-14 standup.m4aand2026-11-14 standup.srt. - Open daytapes.com/app.
- Choose Upload, and add the audio with the transcript file.
- When the page appears, play a minute of the audio and compare the words. Look at names first.
- Rename speakers, then share or hide parts as you like.
If you'd rather send things another way, see email audio to Daytapes and the watch folder. To move recordings over from another service, see export from Otter.
Quick troubleshooting
- Words look right but timing drifts. Try .srt or .vtt instead of .txt, and make sure both files come from the same recording.
- Names are wrong everywhere. That's a mishearing problem. Fix the speaker names in Daytapes and check the transcript for the same names.
- The file is huge. Whisper's output files are small. If the audio is too big to email, upload it.
- The language is wrong. Set the language in your Whisper command. The command line has a
--languageoption.
Mind consent. Recording laws differ by place, and this is not legal advice. Check your state's rules, such as the Reporters Committee's recording guide, and see keeping recordings private before you share a transcript.
Questions
Which Whisper file should I attach to Daytapes?
Use .srt or .vtt if you have them. They carry timing. Daytapes also accepts .txt and .json transcript files, attached with the audio.
Does Whisper tell me who is talking?
The open-source Whisper command line produces text with timing, not speaker names. whisper.cpp lists speaker-turn markers as an experimental option. Daytapes works out who's talking and labels speakers on the page.
Is Whisper free?
The open-source Whisper code and model weights use the MIT license (per the project's GitHub page, read November 14, 2026). You pay with your own computer's time and power. OpenAI's hosted transcription API is a paid service with its own rules.
How big can a file be for OpenAI's API?
OpenAI's speech-to-text guide says files can be up to 25 MB (read November 14, 2026). Daytapes upload accepts recordings up to four hours.
Do I need to transcribe locally at all?
No. Daytapes transcribes audio on its own. A transcript you attach is optional. Local transcription makes sense when audio shouldn't leave your machine or when you already have a good transcript.
Sources
- OpenAI Whisper on GitHub (accessed November 14, 2026)
- OpenAI Whisper source: transcribe.py (command line options) (accessed November 14, 2026)
- whisper.cpp on GitHub (accessed November 14, 2026)
- OpenAI API docs: Speech to text (accessed November 14, 2026)
- OpenAI API docs: Your data (accessed November 14, 2026)
- W3C: WebVTT, Candidate Recommendation Draft (20 May 2026) (accessed November 14, 2026)