Pawfect Pitch

Worried about your big speech? Pawfect Pitch unleashes your potential with AI‐driven speech coaching. Upload audio, get bark‐tastic feedback, and perfect every pitch—no more ruff presentations!

Comment

Inspiration

We noticed how nerve‐racking it is to practice a pitch alone—no real feedback, just endless repeats in front of a mirror. It’s even tougher for people like my friend, who struggles with dyslexia and faces enormous anxiety about speaking in public. Our love of dogs and passion for AI inspired Pawfect Pitch: a fun, friendly, and accessible speech coach that retrieves honest, data‐driven insights on your presentations. By making feedback quick, clear, and supportive, we hope to help everyone—from nervous first‐timers to those with learning challenges—deliver a truly pawfect pitch with confidence.

What it does

Pawfect Pitch is your friendly AI speech coach. Just upload your audio or video, and we’ll give you clear, helpful feedback on everything from pacing and intonation to filler words. Then we suggest easy ways to improve so that you can boost your clarity and confidence—no more second‐guessing. Whether you’re a rookie presenter or a seasoned pro, Pawfect Pitch has your back to help you nail every presentation to the tea.

How we built it

The front end is a simple React/Vite/TypeScript web app with Tailwind for styling. It makes HTTP requests to a Python FastAPI server, which processes and analyzes audio clips with several different libraries and AI models. SciPy and Librosa are used to preprocess the audio clips, and OpenAI's Whisper is used to create a transcript. Numpy is used to compute volume and pace. The ehcalabres/wav2vec2-lg-xlsr-en-speech-emotion-recognition model is used to analyze emotions from audio segments. Last but definitely not least, the HuggingFaceTB SmolLM2-360M-instruct large language model is used to generate summaries and actionable feedback based on transcriptions, we also experimented with larger parameter variants, as well as Qwen 2.5 instruct models.

Challenges we ran into

One of our first problems was getting libraries to properly install, as many of them did not support Python 3.13, only Python 3.11 and below, this took us a bit of time to figure out. We also had trouble with asynchronous Python, with getting files to write and serving HTTP requests at the same time. There were also issues with getting some AI models to work, and VRAM/hardware limitations with some LLM models. We initially also had a lot of trouble coming up with an idea. One of our designers forgot their mouse, and this led to a bit of lost time for us. We also ran into the classic issues with CSS and aligning divs.

Accomplishments that we're proud of

Built a working AI prototype in record time: Despite the typical pressures of a hackathon, we successfully put together a speech recognition and feedback system from scratch.

Delivered in‐depth voice analysis: We’re proud of how our AI can spot filler words, pick up on pacing issues, and provide specific tips to help presenters improve.

Focused on accessibility: Ensuring it helps people with dyslexia or similar challenges was a top priority. Our solution aims to be inclusive for every speaker.

Polished user experience: We didn’t just settle for a functional demo—our clean UI and fun “paw” theme makes the whole process more engaging and less intimidating. In addition, we're super proud of our loading screen, we think something like this sets us apart due to its intuitive design and simple structure.

BONUS Accomplishment–first-time hacker: One of our teammates is a first-time hacker, and he received a ton of experience in both design and dev. He came in knowing next to nothing about the MVP process and was able to gain valuable insights to take forward in future hackathons

What we learned

One of our team members learned to use FastAPI and how to use large language models through hugging face, as well as a plethora of other AI models and machine learning techniques for analyzing audio and speech. It was two of our team members first time using React, they learned how to do conditional rendering, and also learned how to do styling with Tailwind CSS. Our other team member gained some skills in working with Python, as well as more familiarity with Figma.

What's next for Pawfect Pitch

The next steps for Pawfect Pitch involve expanding its capabilities and reach. We plan to integrate video analytics, so it can assess body language and facial expressions in tandem with speech. To make the platform even more accessible, we’ll continue refining our user interface and implementing more robust, customizable feedback options—especially for individuals who speak English as a second language or who have unique learning needs. We also envision introducing collaborative features, allowing small teams to share and review each other’s presentations, and turning Pawfect Pitch into a supportive community as much as an AI coach.

Built With

Share this project:

Updates