Local Transcription Guide

Why Local?

Cloud transcription services charge per minute:

• Google Speech-to-Text: ~$0.006/15 seconds = $1.44/hour
• AWS Transcribe: ~$0.024/minute = $1.44/hour
• OpenAI Whisper API: $0.006/minute = $0.36/hour

At scale (10K+ hours), this becomes significant:

• 10,000 hours × $0.36 = $3,600 (API)
• 10,000 hours × $0 = $0 (local)

Local transcription also means:

• No data leaving your machine
• No rate limits
• Works offline
• One-time hardware investment

Hardware Requirements

Minimum (CPU-only)

• Any modern CPU
• 8GB RAM
• Whisper "tiny" or "base" model
• Speed: ~0.5x real-time (2 hours to transcribe 1 hour)

Recommended (GPU)

• NVIDIA GPU with 6GB+ VRAM
• 16GB RAM
• Whisper "medium" or "large" model
• Speed: 10-30x real-time (1 hour transcribed in 2-6 minutes)

Apple Silicon

• M1/M2/M3 Mac
• 16GB unified memory
• Uses Metal acceleration
• Speed: 5-15x real-time

Setup: Whisper

Installation

# Basic
pip install openai-whisper

# With GPU support (NVIDIA)
pip install openai-whisper torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Basic Usage

# Transcribe single file
whisper audio.mp3 --model medium

# Specify output format
whisper audio.mp3 --model medium --output_format json

# Specify language (faster if known)
whisper audio.mp3 --model medium --language en

Batch Processing

#!/bin/bash
# transcribe_all.sh

for file in ./audio/*.mp3; do
  echo "Processing: $file"
  whisper "$file" --model medium --output_dir ./transcripts
done

Setup: Faster-Whisper

4x faster than standard Whisper with same accuracy.

pip install faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel("medium", device="cuda", compute_type="float16")
segments, info = model.transcribe("audio.mp3")

for segment in segments:
    print(f"[{segment.start:.2f} -> {segment.end:.2f}] {segment.text}")

Setup: MacWhisper (macOS GUI)

1. Download from goodsnooze.gumroad.com/l/macwhisper
2. Drag to Applications
3. Drop audio files to transcribe
4. Uses Apple Silicon acceleration

Good for:

• One-off transcriptions
• Non-technical users
• Quick manual processing

Model Selection

Recommendation: Start with "medium". Use "large" only if accuracy is critical.

Optimization Tips

1. Audio Preprocessing

# Convert to 16kHz mono (optimal for Whisper)
ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav

2. Parallel Processing

from concurrent.futures import ProcessPoolExecutor
import whisper

model = whisper.load_model("medium")

def transcribe(file):
    return whisper.transcribe(model, file)

with ProcessPoolExecutor(max_workers=4) as executor:
    results = executor.map(transcribe, audio_files)

3. GPU Memory Management

import torch
torch.cuda.empty_cache()  # Clear GPU memory between files

4. Skip Already Processed

import os

for audio_file in audio_files:
    transcript_path = audio_file.replace('.mp3', '.json')
    if os.path.exists(transcript_path):
        continue  # Skip already processed
    # ... transcribe

Common Issues

Out of Memory (GPU)

• Use smaller model
• Process shorter segments
• Clear cache between files

Slow Processing

• Check GPU is being used: nvidia-smi
• Use faster-whisper instead
• Reduce audio quality (16kHz mono is sufficient)

Poor Accuracy

• Use larger model
• Specify language if known
• Check audio quality (noise, multiple speakers)

Output Formats

JSON (Recommended)

{
  "text": "Full transcript...",
  "segments": [
    {"start": 0.0, "end": 5.2, "text": "..."},
    {"start": 5.2, "end": 10.1, "text": "..."}
  ],
  "language": "en"
}

SRT (Subtitles)

1
00:00:00,000 --> 00:00:05,200
Welcome to the tutorial

2
00:00:05,200 --> 00:00:10,100
Today we'll cover...

VTT (Web Video)

WEBVTT

00:00:00.000 --> 00:00:05.200
Welcome to the tutorial

00:00:05.200 --> 00:00:10.100
Today we'll cover...

Contribute your setup, benchmarks, or optimization tips.