LLM Transcribe

LLM Transcribe is a Python CLI tool that leverages state-of-the-art Large Language Models to convert audio files into richly annotated transcriptions with timestamps, speaker identification, and emotional context.

Built for developers and technical users who need high-quality transcription with minimal setup, it supports all major LLM providers (Google Gemini, OpenAI, Anthropic) through a single interface, handles arbitrarily long audio files—from quick voice notes to multi-hour meetings—and produces structured output that captures not just what was said but how it was said, including non-verbal cues like laughter, hesitation, or frustration.

Installation

We recommend using uv

uv tool install llm-transcribe

After installation, the llm-transcribe command will be available in your PATH.

Usage

# Simplest way, result will be in audio.txt
llm-transcribe audio.wav

# Specify model and output filename
llm-transcribe -m gemini-2.5-flash -o audio.out.txt audio.wav

# Test setup without processing audio
llm-transcribe audio.wav --test

# Add context for better transcription
llm-transcribe -c "Meeting between John and Kate about Q3 revenue projections" audio.wav

# Get help
llm-transcribe --help

How it Works

1. Audio Processing: Audio file is split into 10-minute chunks with 1-minute overlap using a sliding-window approach.

   • Chunks are technically up to 11 minutes long for context preservation
   • Timing: [0:00-10:00], [9:00-20:00], [19:00-30:00], etc.
   • Supports all common audio formats (WAV, MP3, M4A, etc.)

2. LLM Transcription: Each chunk is transcribed using modern LLMs instead of traditional speech-to-text methods.

   • Audio converted to base64 and sent with carefully crafted prompts
   • Context from previous chunk's final minute maintains speaker consistency
   • Superior accuracy through understanding context and capturing emotional nuances
   • Output format: timestamped lines with speaker identification and non-verbal cues

3. Context Management: Maintains conversation flow across chunks.

   • Previous chunk's final minute provides context for the next chunk
   • Speaker names and roles preserved throughout entire transcription
   • Handles complex timestamp arithmetic (chunk-relative to absolute time)

4. Production Ready: Built on modern Python stack with robust error handling.

   • Retry logic for API failures and rate limiting
   • Cost tracking and provider abstraction via LiteLLM
   • Supports all major LLM providers through unified interface

Output Format

The transcription output includes timestamps and speaker identification:

[01:23:45] John: We should start the meeting now.
[01:23:52] Sarah: I agree. Let me pull up the agenda.
[01:24:05] Speaker 1: Can everyone hear me clearly?
[01:24:08] John: Yes, you're coming through fine.

Speaker Identification Priority

1. Named speakers: When identifiable by name (John, Sarah, etc.)
2. Role-based: When identifiable by role (Manager, Salesperson, etc.)
3. Numerical fallback: Unknown speakers labeled as "Speaker 1", "Speaker 2", etc.

Supported Models

Works with any LLM provider supported by LiteLLM:

• Gemini: gemini-2.5-flash, gemini-1.5-pro
• OpenAI: gpt-4o, gpt-4o-mini
• Anthropic: claude-3-5-sonnet, claude-3-5-haiku
• And many more: See LiteLLM documentation

Requirements

• Python 3.13
• FFmpeg (for audio processing)
• API key for your chosen LLM provider

Configuration

Set your API key as an environment variable:

# For Gemini
export GOOGLE_API_KEY=your_api_key_here

# For OpenAI
export OPENAI_API_KEY=your_api_key_here

# For Anthropic
export ANTHROPIC_API_KEY=your_api_key_here