GPU Compute / Recipe

Audio and transcription jobs

Run Whisper batch transcription or audio processing pipelines as one-off GPU jobs. Billing stops when the script exits — no persistent endpoint needed for occasional workloads.

Prerequisites: npm install -g badgr-cli then badgr login. Setup guide →

Shortcut: badgr transcribe

For a single file or URL, badgr transcribe handles everything — no script needed. Accepts local files ≤50 MB, public URLs, or S3/GCS URIs:

# Local file
badgr transcribe audio.mp3

# Public URL
badgr transcribe https://example.com/recording.mp3

# S3 URI
badgr transcribe s3://my-bucket/audio/interview.wav

Outputs transcript to stdout. Use --output txt or --output json to change the output format. Use badgr run below for batch jobs or custom pipelines.

Batch transcription with Whisper

Run a Python script that transcribes a folder of audio files. Point Badgr at your project folder — it uploads your script and audio files, and mounts them at /app:

badgr run . --cmd "python transcribe.py" --gpu RTX_4090 \
  --env WHISPER_MODEL=large-v3 \
  --env AUDIO_DIR=/app/audio \
  --env OUTPUT_FILE=/app/transcripts.json \
  --max-cost 5

A typical transcribe.py using faster-whisper:

import os, json
from pathlib import Path
from faster_whisper import WhisperModel

model = WhisperModel(
    os.environ.get("WHISPER_MODEL", "large-v3"),
    device="cuda",
    compute_type="float16",
)

audio_dir = Path(os.environ.get("AUDIO_DIR", "/app/audio"))
results = []

for path in audio_dir.glob("*.mp3"):
    segments, info = model.transcribe(str(path))
    text = " ".join(seg.text for seg in segments)
    results.append({"file": path.name, "language": info.language, "text": text})
    print(f"Done: {path.name}")

output = os.environ.get("OUTPUT_FILE", "/app/transcripts.json")
with open(output, "w") as f:
    json.dump(results, f, indent=2)
print(f"Wrote {len(results)} transcripts to {output}")

Custom image (escape hatch)

For unusual setups, bring a container that already has faster-whisper and CUDA configured. Pass model size and paths via --env:

badgr run \
  --image ghcr.io/my-org/whisper-batch:latest \
  --gpu RTX_4090 \
  --env WHISPER_MODEL=large-v3 \
  --env MANIFEST_URL=https://storage.example.com/manifest.json \
  --env OUTPUT_BUCKET=my-transcripts \
  --max-cost 5

Detach and check logs later

Use --detach for large batches so you don't need to keep your terminal open:

# Launch and return immediately
badgr run . --cmd "python transcribe.py" --gpu RTX_4090 \
  --env WHISPER_MODEL=large-v3 \
  --max-cost 5 \
  --detach

# Check progress when ready
badgr logs <deployment-id>

Options

--gpu RTX_4090Recommended for Whisper — fits large-v3 at 24 GB VRAM

--cmd <command>Command to run inside the uploaded project folder.

--image <ref>Custom container with faster-whisper or whisper.cpp — bypasses the runner.

--env KEY=VALUESet environment variables (model size, paths, etc.) — repeatable

--detachReturn job ID immediately, don't stream logs

--max-runtime <min>Auto-stop after N minutes (recommended to cap spend)

--max-cost <$>Auto-stop when spend reaches this amount

Next steps