Voice AI Customer Service

An experimental project exploring how voice models combined with fast, advanced LLMs can replace traditional customer service. The goal is to create AI agents that understand context better and answer customer questions more accurately than human representatives.

Features

• LLM Understanding: Google Gemini 3 Flash (gemini-3-flash-preview) processes customer queries with context awareness
• Text-to-Speech: OpenAI gpt-4o-mini-tts or ElevenLabs converts responses to natural speech
• Customer Service Mode: --gemini-query flag pipes input through LLM before voice synthesis
• Multi-format Input: Supports Markdown, TXT, PDF, and DOCX files

Setup

1. Install dependencies:

   python3 -m pip install -r requirements.txt

   On Linux you may need python3-tk for the GUI picker.

2. Copy and configure environment:

   cp .env.local.example .env.local

3. Edit .env.local with your API keys:

   • GEMINI_API_KEY - For LLM understanding (Get key)
   • OPENAI_API_KEY - For OpenAI TTS
   • ELEVENLABS_API - For ElevenLabs TTS

Usage

Customer Service Mode (LLM + Voice)

Process a customer query through Gemini, then convert the response to speech:

python3 tts.py --gemini-query --text "How do I reset my password?"

Standard Text-to-Speech

python3 tts.py --text "Your order has been shipped" --output notification.mp3

From File

python3 tts.py --input-file response.txt --output ./dist/response.wav --format wav

Test All Connections

python3 tts.py --test

Command-Line Options

Option	Description
--text	Text string to convert to speech
--input-file	Path to a text file (MD/TXT/PDF/DOCX)
--output	Output file path (defaults to ~/Downloads/tts-output.mp3)
--format	Audio format: mp3 or wav (default: mp3)
--provider	TTS provider: openai or elevenlabs
--gemini-query	Process input through Gemini LLM first (customer service mode)
--gemini-api-key	Provide Gemini API key directly
--voice	Override the default voice
--model	Override the model name
--api-key	Provide TTS API key directly
--project	Project ID for OpenAI or ElevenLabs
--choose-file	Open GUI file picker
--chunk-size	Override chunking threshold (default: 3400 chars)
--test	Test API connections

Environment Variables

Gemini (LLM Understanding)

Variable	Description
GEMINI_API_KEY	Google AI API key
GEMINI_MODEL	Model ID (default: gemini-3-flash-preview)

OpenAI (TTS)

Variable	Description
OPENAI_API_KEY	OpenAI API key (required for OpenAI TTS)
OPENAI_MODEL	TTS model (default: gpt-4o-mini-tts)
OPENAI_VOICE	Voice ID (default: alloy)
OPENAI_PROJECT	Project ID for project-scoped keys

ElevenLabs (TTS)

Variable	Description
ELEVENLABS_API	ElevenLabs API key
ELEVENLABS_MODEL	Model (default: eleven_multilingual_v2)
ELEVENLABS_VOICE	Voice ID
ELEVENLABS_STABILITY	Voice stability (0-1)
ELEVENLABS_SIMILARITY	Similarity boost (0-1)
ELEVENLABS_STYLE	Style (0-1)
ELEVENLABS_SPEAKER_BOOST	Speaker boost (true/false)

How It Works

1. Input: Customer query via text, file, or GUI picker
2. LLM Processing (optional): Gemini 3 Flash analyzes the query and generates a helpful response
3. Voice Synthesis: OpenAI or ElevenLabs converts the response to natural speech
4. Output: MP3/WAV audio file ready for playback

The system uses a customer service-optimized prompt that instructs the LLM to be helpful, accurate, and empathetic. The TTS voice is configured for warm, clear delivery suitable for customer interactions.

Supported Input Formats

• *.md / *.txt - Read as UTF-8 (Markdown headings stripped for clean narration)
• *.pdf - Requires PyPDF2
• *.docx - Requires python-docx

References

• Gemini API Docs
• OpenAI TTS Docs
• ElevenLabs Docs