Environment Variable Configuration

March 27, 2026 ยท View on GitHub

This project uses environment variables to configure API keys, model paths, and other parameters. You can create a .env file in the project root directory, and the program will load it automatically.

LLM Evaluation (Required for LLM Fallback Evaluation)

# OpenAI-compatible API
OPENAI_API_KEY=sk-xxx
OPENAI_API_BASE=https://api.openai.com/v1/chat/completions

# Multi-key rotation (configured by index)
OPENAI_API_KEY_1=sk-xxx
OPENAI_API_KEY_2=sk-yyy

When both rule matching and Sentence Transformer semantic matching fail to determine a result, the system automatically calls the LLM API for fallback evaluation.

Gemini API (Optional, only needed when using the gemini_omni model)

GEMINI_API_URL=https://your-api-gateway/v1/chat/completions
GEMINI_API_KEY=your-api-key

HuggingFace (Optional, needed for uploading/downloading datasets)

HF_TOKEN=hf_xxx

Sentence Transformer (Optional, for MQA Semantic Matching)

By default, Qwen/Qwen3-Embedding-0.6B is downloaded from HuggingFace Hub. To specify a local path:

SENTENCE_TRANSFORMER_MODEL=/path/to/your/model

# Model paths
MODEL_PATH=/path/to/modeling_qwen3_new_add_chunk
PROCESSOR_PATH=/path/to/modeling_minicpm26o_clean
MODELING_PATH=/path/to/modeling

# Checkpoint paths
CHECKPOINT_PATH=/path/to/your_checkpoint.pt
DUPLEX_CHECKPOINT=/path/to/duplex_checkpoint.pt
TTS_CHECKPOINT=/path/to/tts_checkpoint.pt

# TTS reference audio (voice cloning)
TTS_REF_AUDIO=/path/to/reference_audio.wav

# Token2wav
TOKEN2WAV_PATH=/path/to/Step-Audio2-main
TOKEN2WAV_MODEL_DIR=/path/to/token2wav

.env File Example

Create a .env file in the project root directory:

# .env
OPENAI_API_KEY=sk-your-key-here
OPENAI_API_BASE=https://api.openai.com/v1/chat/completions

The program searches for .env files in the following order:

  1. Project root directory .env
  2. Current working directory .env
  3. Current working directory .env.local

Existing environment variables will not be overridden by the .env file.