MixedVoices: Analytics and Evals for Voice Agents

MixedVoices is an analytics and evaluation tool for voice agents. Track, visualize, and optimize agent performance through conversation analysis, call quality metrics and call flow charts. Run simulations to test the agent before pushing to production.

Demo

Features

Core Capabilities

• 🌐 Effortless Integration: Python API designed for quick integration, get started in minutes
• 🖥️ Interactive Dashboard: User-friendly interface for all operations
• 📊 Call Flow Analysis: Interactive flowcharts showing conversation paths, patterns and success rates
• 🔄 Version Control: Track and compare agent behavior across different iterations
• 🎯 ML Performance Metrics: Track hallucinations, call scheduling, conciseness, and empathy scores
• 📱 Call Quality Analysis: Monitor interruptions, latency, signal-to-noise ratio, and words per minute
• 🧪 Agent Evaluation: Test and validate agent performance through simulations and stress testing

Installation

pip install mixedvoices

Quick Start

Configure

Before using MixedVoices, configure the models used for various operations by using mixedvoices config.

By default all analytics and transcription happens using OpenAI models. Currently analytics supports all OpenAI GPT models from gpt-3.5 onwards. Transcription supports OpenAI whisper and Deepgram nova-2

mixedvoices config

According to the chosen models, set the environment keys: OPENAI_API_KEY, DEEPGRAM_API_KEY (if nova-2 selected for transcription)

Analytics

Using Python API to analyze recordings

import mixedvoices as mv
from mixedvoices.metrics import Metric, empathy # empathy is an inbuilt metric

# binary=>PASS/FAIL, continuous=>0-10
custom_metric = Metric(name="custom", definition="....", scoring="binary") 

# Create or load a project
project = mv.create_project("dental_clinic", metrics=[empathy, custom_metric])
project = mv.load_project("dental_clinic")  # or load existing project

v1 = project.create_version("v1", prompt="You are a ...") # Create a version
v1 = project.load_version("v1") # or load existing version

# Analyze call, this is blocking, takes a few seconds
v1.add_recording("path/to/call.wav")

# non blocking mode in a separate thread, instantaneous
v1.add_recording("path/to/call.wav", blocking=False)

All recordings added go through the following analysis:-

• Transcription with word level timestamps
• Summarization of transcript
• Classification as Successful/Failed (If success criteria set)
• Breakdown into flow steps
• Metric analysis and scoring
• Call quality analysis (noise, words per minute, latency, interruptions)

Evaluation

Evaluate custom agent

import mixedvoices as mv
from typing import Tuple

# Create agent by inheriting from BaseAgent. Must implement respond
class DentalAgent(mv.BaseAgent):
    def __init__(self, model="gpt-4"):
        self.agent = YourAgentImplementation(model=model)

    def respond(self, input_text: str) -> Tuple[str, bool]:
        response = self.agent.get_response(input_text)
        has_conversation_ended = check_conversation_ended(response)
        return response, has_conversation_ended

project = mv.load_project("receptionist")
v1 = project.load_version("v1")

# Generate test cases using multiple sources
test_generator = mv.TestCaseGenerator(v1.prompt)
test_generator.add_from_transcripts([transcript])  # Add from conversation transcripts
test_generator.add_edge_cases(2)  # Add edge cases
test_generator.add_from_descriptions(["A man from New York, in a hurry"]) # Add from descriptions
test_generator.add_from_project(project)  # Add based on project's recordings
test_generator.add_from_version(v1)  # Add based on version's recordings
test_cases = test_generator.generate()

# Create and run evaluator, can use a subset of metrics
evaluator = project.create_evaluator(test_cases, metric_names=["empathy"])
evaluator.run(v1, DentalAgent, agent_starts=False, model="gpt-4o")

Evaluate Bland AI Agent

# same as above, except instead of defining custom agent, can directly use mv.BlandAgent
evaluator.run(v1, mv.BlandAgent, agent_starts=True, auth_token="", pathway_id="", start_node_id="") 

Using Dashboard

Launch the interactive dashboard from the Command Line:

mixedvoices dashboard

Development Setup

git clone https://github.com/MixedVoices/MixedVoices.git
cd MixedVoices
pip install -e ".[dev]"

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.

Roadmap

• Support other APIs and Open Source LLMs
• Team collaboration features
• Voice based evaluation

---

Made with ❤️ by the MixedVoices Team