Skip to content

Whisper Models

Guide to selecting the right transcription model for your hardware and requirements.

Whisper Model Comparison

Model Quality Speed (CPU) RAM Use Case
tiny Basic ~10x realtime 1 GB Testing only
base Good ~5x realtime 2 GB Low-resource servers
small Very good ~2x realtime 3 GB Budget hardware
medium Excellent ~1x realtime 6 GB Good balance
large-v3-turbo Best ~0.5x realtime (CPU) 3 GB Recommended
large-v3 Best ~0.3x realtime (CPU) 4 GB Maximum quality
large-v2 Great ~0.3x realtime (CPU) 4 GB Legacy

Recommendation

Use large-v3-turbo for the best speed/quality balance. It's nearly as accurate as large-v3 but significantly faster, with lower memory requirements.

Parakeet Models

For RunPod GPU workers, Parakeet models offer dramatically faster transcription:

Model Languages Speed (GPU) Use Case
parakeet-tdt-0.6b-v3 25 EU languages ~100x realtime Default for RunPod

Info

Parakeet models support 25 European languages including English, German, French, Spanish, Italian, and more. For non-European languages, use Whisper models.

Configuration

Set the model via environment variable or web UI Settings page:

# In .env
WHISPER_MODEL=large-v3-turbo

Backend Selection

Backend Platform Set Via
faster-whisper CPU, NVIDIA GPU WHISPER_BACKEND=faster-whisper
mlx Apple Silicon WHISPER_BACKEND=mlx
auto Auto-detect WHISPER_BACKEND=auto (default)

With auto, cast2md selects mlx on Apple Silicon and faster-whisper everywhere else.

Device and Precision

Setting Options Notes
WHISPER_DEVICE cpu, cuda, auto auto detects CUDA availability
WHISPER_COMPUTE_TYPE int8, float16, float32 int8 uses least memory 
WHISPER_BACKEND=mlx
WHISPER_MODEL=large-v3-turbo

WHISPER_BACKEND=faster-whisper
WHISPER_MODEL=large-v3-turbo
WHISPER_DEVICE=cuda
WHISPER_COMPUTE_TYPE=float16

WHISPER_BACKEND=faster-whisper
WHISPER_MODEL=medium
WHISPER_DEVICE=cuda
WHISPER_COMPUTE_TYPE=int8

WHISPER_BACKEND=faster-whisper
WHISPER_MODEL=base
WHISPER_DEVICE=cpu
WHISPER_COMPUTE_TYPE=int8

Chunked Processing

Long episodes are automatically split into chunks to prevent out-of-memory errors:

Setting Default Description
WHISPER_CHUNK_THRESHOLD_MINUTES 30 Episodes longer than this are chunked
WHISPER_CHUNK_SIZE_MINUTES 30 Size of each chunk

This enables 8GB machines to process multi-hour episodes with large-v3-turbo.

Model Management (RunPod)

RunPod pods can use different models. The model is configured per-pod via the Settings page:

  1. Go to Settings -> RunPod
  2. Under "Manage Transcription Models", add or remove models
  3. Select the default model for new pods

API: GET/POST/DELETE /api/runpod/models

Re-transcription

Episodes can be re-transcribed with a newer or better model:

  • API: POST /api/queue/episodes/{id}/retranscribe
  • Batch: POST /api/queue/batch/feed/{id}/retranscribe (re-transcribes all episodes using an outdated model)

The transcript_model column on each episode tracks which model produced the current transcript.