frisket DocsSearch

Recipes / Recordings

Turn council recordings into checked transcripts

Transcribe two five-minute excerpts of city council meetings, then check the transcript by listening to the recording.

Open the recordings and their source details

The sample opens on the Council audio sheet. Each row is one recording, a five-minute excerpt of an Ann Arbor city council meeting. To listen, click the small waveform icon at the left of a media cell; click it again to stop.

The Council meetings sheet has each meeting's date, a link to the original recording, its license and its author, so you can cite where the audio came from. Click its tab, then click a row to see all of these at once.

Waveforms and filenames in the Council audio sheet
Recording source link in row details

Open Transcribe Compare

Different transcription engines make different mistakes, and you can't tell which is better without hearing them on your own audio. Transcribe Compare runs several engines on the same short passage and shows the results side by side. It's a scratch space: nothing from it is saved to your sheet. On the Media tab, click Transcribe Compare. It opens in its own tab beside your sheets, with an empty drop area.

Media ribbon with Transcribe Compare highlighted
Transcribe Compare with an empty recording drop area

Add the recording

Transcribe Compare works on files from your computer, not on rows in a sheet. Click Download the files at the top of this page if you haven't already. Then click the drop area and choose the Policy Committee recording, or drag it onto the drop area. Its name appears in the list on the left.

Recording

Every engine hears the same file, so differences come from the engines, not the audio.

Set to: ann-arbor-policy-committee-2022-04-19-excerpt.mp3

Policy Committee recording added to the comparison

Choose the engines

Each chip along the top is one engine with its settings. Frisket starts you with a couple, depending on what's set up. You want two: Whisper, which runs where Frisket runs, and Microsoft MAI-Transcribe 2 (OpenRouter), a hosted engine. Comparing them shows whether paying for the hosted one is worth it on your audio.

  1. If either is missing, click + Engine and choose it from the list.
  2. Remove any other engine with the × on its chip.
  3. Click the settings icon on Whisper's chip and check Model size is base.

MAI-Transcribe 2 needs an OpenRouter API key, or Frisket Cloud if your account offers it.

Engines

One engine that runs where Frisket runs, and one hosted engine.

Set to: Whisper (multilingual transcription), base; Microsoft MAI-Transcribe 2 (OpenRouter)

MAI-Transcribe 2 in the engine list
Whisper and MAI-Transcribe 2 selected for comparison
Whisper settings with base model selected

Choose the passage length

First seconds sets how much of the start of the recording each engine hears. It starts at 600, ten minutes, which would send the whole excerpt to every engine. Set it to 30. Thirty seconds is enough to hear how each engine handles names and unclear speech, and keeps the cost of a hosted engine small.

1
First seconds

Only this many seconds are sent to each engine.

Set to: 30

First seconds set to 301

Run the comparison

Click Run comparison. Each engine's column fills in as it finishes. Whisper costs nothing to run. MAI-Transcribe 2 is charged by the provider; for these 30 seconds it cost under a tenth of a cent. If Frisket asks you to approve an estimate, check it and click Run it.

Run comparison for the first 30 seconds

Inspect the comparison

Read the two columns side by side. MAI-Transcribe 2 returns a few long passages and Whisper many short ones, so their timestamps don't line up, and MAI's later rows read "no text". Compare the words, not the rows.

Where the words differ, listen: click Source to open the recording's player, find the passage and play it. Pay most attention to names, numbers and short words that change the meaning. In this excerpt the engines spelled a name differently and heard one phrase differently.

Whisper and MAI transcripts side by side
Audio player for checking the compared passage

Where the engines disagreed

SourceWhisper · baseMAI-Transcribe 2
Policy Committee · a nameJamie ZoneskiJamie Zaneski
Policy Committee · 0:20She's out there todayShe's out ill today

Both are real outputs for the same opening 30 seconds. Listen to these two moments before using either version.

Transcribe the recordings

When you've chosen an engine, transcribe the recordings and save the results on the sheet. Go back to the Council audio sheet. On the Media tab, click Transcribe and set it up as below, then click Run. Whisper runs where Frisket runs, so there's no provider charge. Both recordings are transcribed, one row each.

Find it under Media › Transcripts › Transcribe

Media ribbon with Transcribe highlighted
1
Media

The column holding the recordings.

Set to: media

2
Engine

Runs where Frisket runs: on your computer with Frisket Desktop.

Set to: Whisper (multilingual transcription)

3
Model size

A small model that handles clear speech. Larger sizes are slower.

Set to: base

Transcribe input column, engine and model size123
4
Transcript

The column for the full text.

Set to: transcript

5
Timestamped segments

The column for the text split into timed passages. Document view needs it to play the recording from a passage.

Set to: transcript_segments

Transcript and timestamped segment output columns45

Read the transcript while listening

Document view shows a recording's player beside its transcript, split into short timed passages.

  1. Click Document above the grid.
  2. Choose a recording from the list on the left.
  3. Click a passage to play the recording from there.

Before you quote anyone, play the passage and check the names, numbers and wording against what you hear.

Document view above the Council audio sheet
The recording list beside the selected audio and transcript
Timestamped transcript with a passage selected
The saved transcript and timestamped segment columns

You should end up with

Three new columns on the Council audio sheet: transcript, transcript_segments and detected_language.

filename transcript transcript_segments detected_language
ann-arbor-policy-committee-2022-04-19-excerpt.mp3 Andy Bouchek and Mary Beth McGowan, but the three of us… 43 transcript_segments en
ann-arbor-icpoc-2020-07-01-excerpt.mp3 We had sworn on the… 70 transcript_segments en

Transcription can differ between runs, even with the same model.

If it doesn't match

Transcribe doesn't offer the media column. Check you're on the Council audio sheet. The Council meetings sheet has links to the recordings, not the recordings themselves.

MAI-Transcribe 2 shows a setup message. Add an OpenRouter API key or connect Frisket Cloud, or compare Whisper with another engine you have set up.

The comparison ran on the whole five minutes. First seconds was still 600. Set it to 30 before clicking Run comparison.

Document view shows plain text with no timestamps. The Timestamped segments column is missing. Run Transcribe again with it named.