Extracting a Recipe from YouTube with AWS Transcribe

I watched a DW Food video about Finnish blueberries and wanted to extract a recipe mentioned midway through. The recipe was buried in 10 minutes of footage with no timestamp and no pinned comment. Rather than rewinding and writing it down, I built a bash script that pulls the audio, sends it through Amazon Transcribe, and outputs a clean text file. The recipe itself — a Finnish blueberry pie adapted for UK ingredients — is documented on notes.digitalden.cloud.

---

Why a Transcript

I am new to baking. Having the recipe as plain text meant I could paste it into AI and work with it properly — asking what each ingredient is, whether UK substitutes exist, converting to grams, and clarifying what each step actually does. That back and forth is hard to do with a video. With a transcript it is straightforward.

---

Pipeline Overview

graph LR
    A[YouTube Video] -->|yt-dlp| B[MP3 File]
    B -->|aws s3 cp| C[S3 Bucket]
    C -->|start-transcription-job| D[Amazon Transcribe]
    D -->|JSON output| C
    C -->|python3 extract| E[transcript.txt]

Step	Tool	Action
1	yt-dlp	Extract audio from YouTube as MP3
2	AWS CLI	Upload MP3 to S3
3	Amazon Transcribe	Start transcription job
4	AWS CLI	Poll for job completion
5	AWS CLI + Python	Download JSON, extract plain text

---

Step 1: Extract Audio

yt-dlp -x --audio-format mp3 "https://www.youtube.com/watch?v=57HYrItZiKc"

This downloads the video, extracts the audio track, and converts it to MP3 in one command. Output: Why Finns are obsessed with blueberries.mp3.

What is yt-dlp
A command line tool for downloading video and audio from YouTube and most other video platforms. The -x flag extracts audio only and --audio-format mp3 converts it directly, so you never download the full video file.

---

Step 2: The Transcribe Script

The bash script handles S3 upload, job creation, polling, and text extraction in a single run.

#!/bin/bash
BUCKET="your-bucket-name"
REGION="eu-west-1"

if [ -z "$1" ]; then
  echo "Usage: ./transcribe.sh <audio-file.mp3>"
  exit 1
fi

AUDIO_FILE="$1"
FILENAME=$(basename "$AUDIO_FILE" .mp3)
JOB_NAME="${FILENAME// /-}-$(date +%s)"
S3_KEY="transcribe-input/${FILENAME}.mp3"
OUTPUT_KEY="transcribe-output/${JOB_NAME}.json"
OUTPUT_TXT="${FILENAME}.txt"

echo "Uploading: $AUDIO_FILE"
aws s3 cp "$AUDIO_FILE" "s3://$BUCKET/$S3_KEY" --region "$REGION"

echo "Starting Transcribe job: $JOB_NAME"
aws transcribe start-transcription-job \
  --transcription-job-name "$JOB_NAME" \
  --media "MediaFileUri=s3://$BUCKET/$S3_KEY" \
  --media-format mp3 \
  --language-code en-US \
  --output-bucket-name "$BUCKET" \
  --output-key "$OUTPUT_KEY" \
  --region "$REGION"

echo "Polling for completion..."
while true; do
  STATUS=$(aws transcribe get-transcription-job \
    --transcription-job-name "$JOB_NAME" \
    --region "$REGION" \
    --query "TranscriptionJob.TranscriptionJobStatus" \
    --output text)

  echo "Status: $STATUS"
  [ "$STATUS" = "COMPLETED" ] && break
  [ "$STATUS" = "FAILED" ] && echo "Job failed." && exit 1
  sleep 15
done

echo "Downloading transcript..."
aws s3 cp "s3://$BUCKET/$OUTPUT_KEY" "/tmp/${JOB_NAME}.json" --region "$REGION"

echo "Extracting plain text to $OUTPUT_TXT"
python3 -c "
import json
with open('/tmp/${JOB_NAME}.json') as f:
    data = json.load(f)
print(data['results']['transcripts'][0]['transcript'])
" > "$OUTPUT_TXT"

echo "Done. $OUTPUT_TXT"

Usage

chmod +x transcribe.sh
./transcribe.sh "Why Finns are obsessed with blueberries.mp3"

---

Design Decisions

S3 output path

The script uses --output-bucket-name and --output-key to write the JSON directly to a known S3 path. Without these flags, Transcribe writes to a Transcribe-managed URI that requires an extra API call to resolve. This keeps the download step to one aws s3 cp with a predictable path.

--output-bucket-name "$BUCKET" \
--output-key "$OUTPUT_KEY"

Job name collisions

JOB_NAME="${FILENAME// /-}-$(date +%s)"

A Unix timestamp is appended so reruns of the same file never collide. Transcribe rejects duplicate job names.

Text extraction

data['results']['transcripts'][0]['transcript']

The Transcribe JSON output contains word-level timings, confidence scores, and speaker labels. For this use case none of that matters. I only need the full transcript as a single text block.

Polling interval

sleep 15

A 10-minute audio file typically completes in 2–3 minutes. Polling every 15 seconds keeps it responsive without excessive API calls.

---

Output Quality

The transcript came back clean:

What worked

Content	Result
English narration	Clean, accurate
Place names (Helsinki, Finland)	Correct
Cooking terms (rhubarb, blueberry, oats)	Correct
Numbers and measurements	Correct

Where it struggled

Content	Expected	Output
mustikka (Finnish for blueberry)	mustikka	Musica
Varpu (Finnish name)	Varpu	Varpoo

This is expected. The job used en-US as the language code. The model is not trained on Finnish proper nouns. For a primarily English video with occasional Finnish words, the accuracy was more than enough to work with.

Language code selection
Amazon Transcribe supports Finnish (fi-FI). If the source audio were predominantly Finnish, switching the language code would improve accuracy on Finnish words but degrade English sections. For mixed-language content, the dominant language is usually the better choice.

---

S3 Bucket Structure

After a run, the bucket looks like this:

s3://your-bucket-name/
├── transcribe-input/
│   └── Why Finns are obsessed with blueberries.mp3
└── transcribe-output/
    └── Why-Finns-are-obsessed-with-blueberries-1740153600.json

Input and output are separated by prefix. The JSON filename includes the timestamp so multiple runs are easy to trace.

---

Cost

Resource	Cost
Amazon Transcribe	~$0.24 per 10 minutes of audio
S3 storage	Negligible for a single MP3 + JSON
S3 requests	Negligible

For a one-off extraction this costs almost nothing.

---

What Comes Next

The CLI version works but requires a terminal and manual execution. The next step is making it event-driven.

graph LR
    A[MP3 → S3] --> B[Lambda: Start Transcribe]
    B --> C[EventBridge: Job Complete]
    C --> D[Lambda: Extract + Bedrock]
    D --> E[Markdown → S3]

Component	Role
S3 event notification	Triggers on MP3 upload
Lambda (start)	Creates Transcribe job
EventBridge rule	Catches TranscribeJobStateChange with COMPLETED status
Lambda (process)	Downloads transcript, sends to Claude via Bedrock, formats output
S3 (output)	Stores final markdown file

Drop an MP3 into S3 and a formatted markdown file appears. No polling, no terminal. Markdown is the output format because all my sites run on Jekyll with the Chirpy theme, so every post is a .md file with YAML front matter. A pipeline that outputs markdown means the transcript lands in a format I can publish directly. I will document that build when I get there.

For now the CLI version is enough. And the Finnish blueberry pie it extracted was 10/10.

---

Documented February 2026.