local-media-transcription

Local Media Transcription

A skill that transcribes local audio/video files and, as needed, organizes them into minutes, action items, and PPT key points.

When to Use

• Transcription, transcription, transcribe mp4, meeting transcript
• Transcribe from mp4, Minutes from recordings, Transcribe with whisper
• When you want to extract key points from Teams recordings, Screenpresso recordings, or meeting audio
• When you want to complete everything in a local environment

Supported Inputs

• MP4 / M4A / MP3 / WAV / WEBM
• Japanese meetings, internal meetings, customer meeting recordings

Outputs

• Raw transcription .txt
• Transcription with speaker diarization .txt / .md
• Minutes .md
• Customer-facing cleaned-up minutes .md
• Action Item .md
• Key points for PowerPoint .md

Workflow

1. Input check

   • Check whether the target file exists
   • Check the length, size, and modification date/time

2. Tool check

   • ffmpeg, ffprobe, whisper Check whether the CLI is available
   • If unavailable, stop the local transcription and present alternatives

3. Media check

   • ffprobe Obtain duration with
   • If it is a long recording, estimate the processing time

4. Execute transcription

   • For Japanese meetings, use the following as the basic form

   $env:PYTHONIOENCODING='utf-8'
   whisper "<media-file>" --language Japanese --task transcribe --model turbo --fp16 False --verbose False --output_format txt --output_dir "<output-dir>"
   

5. Quality check

   • Output .txt Read the first few dozen lines of to check whether it is coherent
   • At minimum, check whether it is readable as speech and whether the file is not empty

6. Speaker diarization as needed

   • If speaker diarization is needed, check the available diarization methods
   • If available, generate a transcription with speaker labels
   • If unavailable, proceed without speaker diarization or stop while clearly stating the constraint

7. Format as needed

   • Minutes / Customer-facing cleaned-up minutes / Action items / PPT key points
   • Do not overstate; correct speech recognition fluctuations and organize

Decision Points

A. When the tool is missing

• ffmpeg / whisper If there is no
• As an alternative, guide them to paste the Teams transcription text or install it in advance

B. When transcription alone is sufficient

• .txt Output and finish

C. When meeting organization is also needed

• Additionally generate the following deliverables
  • Minutes
  • Customer-facing cleaned-up minutes
  • Action Item
  • PPT key points

D. When speaker diarization is needed

• Check the diarization tool or existing environment
• If available, 話者A / 話者B output once with labels such as
• At the stage where names are not confirmed, do not assert estimated names
• If unavailable, clearly state "proceed without speaker diarization / correct manually later"

E. When placing in customer-facing materials

• Visually correct proper nouns and speech recognition errors
• Do not assert ambiguous points definitively; phrase them as "understood to be" or "explained as".
• Remove internal conversations, unconfirmed information, and small talk.
• For customer-facing output, organize it so that "decisions / items under consideration / next actions" are clear at a glance.

Completion Criteria

• Confirmation of the existence of the input media has been completed.
• The transcript .txt has been generated.
• There are no fatal breakdowns in the initial confirmation.
• For requests requiring speaker diarization, either a feasibility judgment or diarized output is explicitly stated.
• If requested, the meeting minutes / customer-facing clean version / actions / PPT key points have been generated.

Quality Rules

• Do not add assertions that are not in the original audio.
• Be aware that numbers and proper nouns may be misrecognized.
• In customer-facing summaries, do not assume measures are a given.
• In customer-facing meeting minutes, do not leave internal discussions or unapproved matters as they are.
• In speaker diarization, do not arbitrarily convert labels that cannot be confirmed as personal names into real names.
• For long audio, prioritize successful transcription first, then proceed to organization.

References

• transcription-workflow.md