Skip to content
This repository was archived by the owner on Jul 12, 2026. It is now read-only.

PROVIDERS.md

Latest commit

 

History

History
133 lines (91 loc) · 3.53 KB

PROVIDERS.md

File metadata and controls

133 lines (91 loc) · 3.53 KB
Raw

Transcription Providers

This guide covers detailed configuration for all supported transcription providers.

Overview

Provider Type API Key Required Best For
Deepgram Nova 3 Cloud Yes Production, accuracy, speaker diarization
ElevenLabs Cloud Yes High-quality voice processing
Whisper Local No Privacy, offline use, 100+ languages

Deepgram

Cloud-based provider with excellent accuracy and full feature support.

Setup

# Get API key from: https://console.deepgram.com/
export DEEPGRAM_API_KEY='your-api-key-here'

# Or add to .env file
echo "DEEPGRAM_API_KEY=your-api-key-here" >> .env

Features

  • Speaker diarization
  • Punctuation and formatting
  • Topic detection
  • Sentiment analysis
  • Real-time streaming

ElevenLabs

Cloud-based provider optimized for voice processing.

Setup

# Get API key from: https://elevenlabs.io/api
export ELEVENLABS_API_KEY='your-api-key-here'

# Or add to .env file
echo "ELEVENLABS_API_KEY=your-api-key-here" >> .env

Whisper (Local)

OpenAI's Whisper runs locally—no API key needed. Supports 100+ languages.

Installation

# Basic installation
uv add openai-whisper torch

# For GPU acceleration (CUDA 11.8)
uv add openai-whisper torch torchaudio --index-url https://download.pytorch.org/whl/cu118

# Verify installation
python -c "import whisper; print('Whisper installed successfully')"

Model Selection

Model Parameters Disk Space RAM Usage VRAM Usage Quality
tiny 39M 75MB ~1GB ~1GB Basic
base 74M 142MB ~1GB ~1GB Good
small 244M 461MB ~2GB ~2GB Better
medium 769M 1.5GB ~5GB ~5GB Great
large 1.5B 2.9GB ~10GB ~10GB Best

Configuration

export WHISPER_MODEL='base'        # Model size: tiny, base, small, medium, large
export WHISPER_DEVICE='cuda'       # Device: cuda or cpu
export WHISPER_COMPUTE_TYPE='float16'  # Precision: float16 or float32

Performance Tips

  • tiny/base: Fast, good for drafts or real-time use
  • small/medium: Balanced accuracy and speed
  • large: Best accuracy, requires significant VRAM
  • Use cuda device when GPU is available for 5-10x speedup

Parakeet (Removed)

Parakeet support is no longer part of this repository. If you need local transcription without API keys, use Whisper.

Provider Selection

The CLI supports automatic provider selection or explicit choice:

# Auto-select based on available API keys/models
audio-extraction-analysis transcribe audio.mp3 --provider auto

# Explicit provider selection
audio-extraction-analysis transcribe audio.mp3 --provider deepgram
audio-extraction-analysis transcribe audio.mp3 --provider whisper

Selection Priority (auto mode)

  1. Deepgram (if DEEPGRAM_API_KEY is set)
  2. ElevenLabs (if ELEVENLABS_API_KEY is set)
  3. Whisper (if installed)

Environment Variables Reference

Variable Provider Description
DEEPGRAM_API_KEY Deepgram API key for Deepgram
ELEVENLABS_API_KEY ElevenLabs API key for ElevenLabs
WHISPER_MODEL Whisper Model size (tiny/base/small/medium/large)
WHISPER_DEVICE Whisper Device (cuda/cpu)
WHISPER_COMPUTE_TYPE Whisper Precision (float16/float32)