Updated Sep 18, 2026

Utilities

Transcribe speech from an audio file or folder

Speech to txt, vtt, srt, and json from an audio file or folder.

Agent

overview

What whisper does

You have an MP3, WAV, or a folder of those files, and the next node needs the spoken words as text. Set --audio for one file or --audio-folder for a directory. The wrapper calls OpenAI Whisper, writes into the output folder, and leaves txt, vtt, srt, tsv, and json for each input. --model picks tiny, base, small, medium, or large. --language pins a code such as en.

A folder input walks common audio and video extensions. --output-format can be txt, vtt, srt, tsv, json, or all. --task translate writes English text from another spoken language. --initial-prompt, --temperature, and --beam-size tune decoding. --word-timestamps adds word-level timing when the next node needs it.

Whisper reads the audio you already have. It does not download a stream or invent a caption track. Pair it after yt-dlp when the enclosure is audio and YouTube captions are missing or blocked. A long episode on the base model can take tens of minutes on a CPU worker. Treat an empty folder as a failed transcribe, not a silent show.

source github.com/openai/whisper

use cases

Where whisper fits

Transcribe one MP3

Set --audio and --model base. The folder holds episode.txt plus the other formats when --output-format is all.

Transcribe a folder of episodes

Pass --audio-folder. Whisper walks the directory and writes one text file per audio file.

Pin English on a known show

Set --language en so the decoder does not guess. Use another ISO code when the show is not English.

Keep only plain text

Set --output-format txt when later nodes do not need vtt or srt.

reference

whisper inputs and flags

11 inputs
NameTypeFlagDescription
audioFILE--audioOne audio file. Use this or --audio-folder.
audio-folderFOLDER--audio-folderFolder of audio files to transcribe.
modelSTRING--modeltiny, base, small, medium, large, large-v2, or large-v3. Default base.
languageSTRING--languageLanguage code such as en. Auto-detects when empty.

Showing key inputs. whisper exposes 11 inputs in total.

Full flag reference (11 inputs)
NameTypeFlagDescription
audioFILE--audioInput audio file (mp3, wav, m4a, flac, etc.)
audio-folderFOLDER--audio-folderFolder containing audio files to transcribe
modelSTRING--modelModel size: tiny, base, small, medium, large, large-v2, large-v3 (default: base)
languageSTRING--languageLanguage code (e.g., en, es, fr, de, ja). Auto-detects if not specified.
taskSTRING--taskTask to perform: transcribe (default) or translate (to English)
output-formatSTRING--output-formatOutput format: txt, vtt, srt, tsv, json, all (default: all)
word-timestampsSTRING--word-timestampsInclude word-level timestamps in output (true/false)
initial-promptSTRING--initial-promptOptional text to provide as initial prompt for the decoder
temperatureSTRING--temperatureSampling temperature (0 = greedy decoding, default: 0)
beam-sizeSTRING--beam-sizeNumber of beams for beam search (default: 5)
verboseSTRING--verbosePrint progress and debug messages (true/false)

example

Run whisper

whisper · command
# one file; last argv is the output folder on Trickestwhisper --audio episode.mp3 --model base --language en --output-format all /hive/out # folder of fileswhisper --audio-folder /hive/in/audio --model base --language en --output-format txt /hive/out
sample output
When I learned that it was okay to not know everything, that it is okay to ask questions and ask for understanding, allow myself to be mentored, that is when opportunities presented themselves. Welcome to the example.com briefing. I am your host.

guidance

Choosing whisper

Use whisper when you already have audio and need speech as text. Use youtube-transcript when a public caption track is enough. Use yt-dlp when you still need to download the media file.

yt-dlp

Downloads the media file. Run it first when the input is still a URL or RSS enclosure.

youtube-transcript

Reads a public caption track. Faster when captions exist and the fleet IP is not blocked.

wget

Downloads HTTP files. It does not transcribe speech.

faq

whisper questions

Set --audio and --model base. Leave --output-format at all to get txt, vtt, srt, tsv, and json in the output folder.

Run whisper yourself

An audio file feeds the whisper node. The folder holds txt, vtt, srt, tsv, and json. The text sample below is illustrative, not a fleet run.

Facts on this page come from the live Trickest tool library.