env.jetson.example

# ============================================================================
# NEMOTRON VOICE AGENT - ENVIRONMENT CONFIGURATION
# ============================================================================
# Copy this file to .env and configure the values for your deployment.
# Variables with values are defaults; commented variables are optional overrides.


# ----------------------------------------------------------------------------
# REQUIRED CREDENTIALS
# ----------------------------------------------------------------------------

# Your NVIDIA API key from https://build.nvidia.com
# Export it as an environment variable in your shell:
#   export NVIDIA_API_KEY=<your-nvidia-api-key>
# Docker Compose will automatically use the environment variable if this is left empty.
NVIDIA_API_KEY=

# Huggingface token needed for LLM model download
# Export it as an environment variable in your shell:
#   export HF_TOKEN=<your-huggingface-token>
# Docker Compose will automatically use the environment variable if this is left empty.
HF_TOKEN=

# ----------------------------------------------------------------------------
# TURN SERVER CREDENTIALS
# ----------------------------------------------------------------------------

TURN_SERVER_URL=
TURN_USERNAME=
TURN_PASSWORD=



# ----------------------------------------------------------------------------
# DOCKER IMAGE CONFIGURATION
# ----------------------------------------------------------------------------

# Version tag for the python-app and ui container image
IMAGE_VERSION=1.0.0-arm64


# ----------------------------------------------------------------------------
# PIPELINE CONFIGURATION
# ----------------------------------------------------------------------------

# Transport mode for the voice agent: "WEBSOCKET" or "WEBRTC"
# WEBSOCKET: Uses FastAPI WebSocket transport (pipeline_websocket.py)
# WEBRTC: Uses WebRTC transport (pipeline_webrtc.py) - Default
TRANSPORT=WEBRTC

# Path to the prompt catalog YAML file containing system prompts
PROMPT_FILE_PATH=./config/prompt.yaml

# Enable speculative speech processing for lower response latency
# When enabled, the bot starts generating responses before user finishes speaking
ENABLE_SPECULATIVE_SPEECH=false

# Voice Activity Detection (VAD) engine: "ASR" (recommended) or "Silero"
VAD_PROFILE=ASR

# Maximum conversation turns to retain in context
# For multilingual or emotion-aware use cases, set the limit to 3-5 for best accuracy
CHAT_HISTORY_LIMIT=20


# Audio dump directory for debugging and analysis
AUDIO_DUMP_PATH=./audio_dumps

# JSON file containing word-to-IPA mappings for pronunciation correction
TTS_IPA_FILE_PATH=./config/ipa.json

# Number of 10ms audio chunks to buffer for output (controls audio latency)
# Default: 5 chunks (50ms buffer) for WebRTC - optimized for low latency
# WebSocket: 10 chunks (100ms buffer) - more stable for network variations
# High Concurrency: 10-40 chunks (100-400ms buffer) - prevents audio glitches under load
AUDIO_OUT_10MS_CHUNKS=5

# Number of workers for HTTP server (handles concurrent connections)
WORKERS=1

# ----------------------------------------------------------------------------
# OPENTELEMETRY TRACING
# ----------------------------------------------------------------------------
ENABLE_TRACING=false
OTEL_CONSOLE_EXPORT=false
# Add Endpoint for OTEL Collector.
# For Phoenix local deployment: `docker run -p 6006:6006 -p 4317:4317 -i -t arizephoenix/phoenix:latest`
# For gRPC (port 4317): Use host:port format (e.g., localhost:4317 or phoenix:4317)
# For HTTP (port 4318): Use http://host:port format (e.g., http://localhost:4318)
#OTEL_EXPORTER_OTLP_ENDPOINT=phoenix:4317

# ----------------------------------------------------------------------------
# ASR (AUTOMATIC SPEECH RECOGNITION) CONFIGURATION
# ----------------------------------------------------------------------------

# ASR endpoint URL
# Example: grpc.nvcf.nvidia.com:443 (cloud) or localhost:50051 (local NIM)
ASR_SERVER_URL=localhost:50051

# ASR model identifier
ASR_MODEL_NAME=parakeet-1.1b-en-US-asr-streaming



# ----------------------------------------------------------------------------
# TTS (TEXT-TO-SPEECH) CONFIGURATION
# ----------------------------------------------------------------------------

# TTS endpoint URL
# Example: grpc.nvcf.nvidia.com:443 (cloud) or localhost:50051 (local NIM)
TTS_SERVER_URL=localhost:50051

# Default voice identifier (format: Model.Language.VoiceName)
TTS_VOICE_ID=Magpie-Multilingual.EN-US.Aria

# TTS model identifier
TTS_MODEL_NAME=magpie_tts_ensemble-Magpie-Multilingual

# Language code for speech synthesis
TTS_LANGUAGE=en-US



# ============================================================================
# LLM (LARGE LANGUAGE MODEL) CONFIGURATION
# ============================================================================

# Models: nvidia/Nemotron-Mini-4B-Instruct, nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8, Qwen/Qwen3-4B-Instruct-2507, RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized-w4a16
NVIDIA_LLM_URL=http://localhost:9000/v1
NVIDIA_LLM_MODEL=RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w4a16

# GPU memory utilization ratio for the LLM model (0.0 to 1.0)
# Controls how much GPU memory the model can use. Lower values leave more memory for other processes.
GPU_MEMORY_UTILIZATION=0.15

SYSTEM_PROMPT_SELECTOR=llama/flowershop