Inspiration
Speech plays an important role in communication, education, interviews, presentations, and public speaking. However, general feedback such as “speak more clearly” or “improve your pace” does not explain exactly where a speaker needs to improve.
We were inspired to build SpeechLens AI, a tool that makes speech feedback more specific, measurable, and actionable. Our goal is to help users compare a reference recording with a test recording and identify the time intervals where their acoustic characteristics differ.
Our guiding idea is simple: Every flaw has a timestamp. Every score has evidence.
What it does
SpeechLens AI is a contrastive speech analytics platform that compares two audio recordings and highlights potentially important differences in speech delivery.
Key features include:
- Reference and test comparison: Analyze two recordings side by side.
- Acoustic feature analysis: Examine characteristics such as pitch, energy, MFCCs, pauses, and speech rhythm.
- Timestamped difference detection: Identify intervals where measured acoustic features differ noticeably.
- Interactive timeline: Explore the intervals highlighted by the analysis.
- Feature visualization: Understand how acoustic measurements change throughout the recordings.
- Interval inspection: Review the evidence behind detected differences instead of relying only on a single score.
- Demo scenarios: Experiment with sample recordings and different speech-delivery patterns.
SpeechLens AI is designed to support speakers, students, presenters, trainers, and researchers who want a more detailed way to examine speech delivery.
How we built it
We built SpeechLens AI using Python and Streamlit, with audio-processing and numerical-analysis libraries.
The workflow begins with a reference recording and a test recording. The audio is processed into measurable acoustic features, including energy, pitch-related measurements, MFCCs, and timing-related characteristics. The system compares these features across time intervals and calculates differences that may indicate changes in delivery.
The results are presented through an interactive Streamlit dashboard, where users can inspect summary information, timelines, individual intervals, and feature-level evidence.
Our main tools and technologies include:
- Python — core application logic and analysis
- Streamlit — interactive web dashboard
- Librosa — audio feature extraction and analysis
- NumPy — numerical computation
- Pandas — organizing analysis results
- Plotly and other visualization tools, where integrated — interactive charts and result exploration
- Git and GitHub — source-code management and collaboration
We organized the project into reusable source modules and included setup instructions and tests to make it easier to run, understand, and extend.
Challenges we ran into
One of our main challenges was turning complex audio signals into understandable feedback. Audio recordings contain many variations, and not every difference necessarily represents a speech problem.
Another challenge was comparing recordings over time. Differences in recording duration, speaking speed, pauses, and acoustic conditions can affect the analysis. We needed to present these differences clearly without making unsupported claims about a speaker's performance.
We also worked to make the dashboard useful for users who may not have a background in signal processing. Visual timelines, feature summaries, and interval-level inspection help make the analysis easier to explore.
These challenges reinforced the importance of transparent measurements, clear explanations, and honest limitations.
Accomplishments that we're proud of
We are proud to have developed a working prototype that brings audio feature analysis and interactive visual feedback into one application.
Our key accomplishments include:
- Building a reference-versus-test speech analysis workflow.
- Presenting acoustic differences through an interactive dashboard.
- Providing timestamped intervals for closer inspection.
- Making feature-level evidence easier to explore.
- Organizing the codebase and publishing the project on GitHub.
- Creating a foundation that can be extended with more advanced speech alignment and machine-learning techniques.
Most importantly, we transformed the idea of general speech feedback into a more inspectable and evidence-oriented analysis experience.
What we learned
Through this project, we learned how audio-processing techniques can be used to extract meaningful acoustic measurements from speech. We also learned the importance of comparing signals over time, visualizing results effectively, structuring a Python application, and documenting a project for other developers.
We learned that a useful AI-powered tool should not simply produce a score. It should also show the evidence behind that score and explain what the results can—and cannot—tell the user.
A measured acoustic difference is not automatically a confirmed speech flaw. SpeechLens AI therefore treats detected differences as signals for further inspection rather than definitive judgments.
What's next for SpeechLens AI
Our next steps are to improve the accuracy and reliability of the comparison pipeline, explore stronger temporal alignment methods, and evaluate the analysis against carefully prepared recordings with known delivery differences.
We also plan to investigate more advanced models for speech representation and contrastive analysis, improve explanations for detected intervals, and expand the range of supported speech scenarios.
Our long-term vision is to make speech analysis more transparent and actionable, helping people understand not only that two recordings differ, but also where they differ, which acoustic measurements changed, and why those changes deserve attention.
Log in or sign up for Devpost to join the conversation.