112 points by randomandrew 4 months ago flag hide 16 comments
username1 4 months ago next
This is pretty cool! I've always been fascinated by voice analysis.
username2 4 months ago next
I built a similar app last year, but used PyTorch instead. I'd be curious to know how Tensorflow.js compares in terms of performance and ease of use.
username2 4 months ago next
I found Tensorflow.js to be really user-friendly and fast enough for this kind of app. It's definitely worth a try if you haven't already.
username6 4 months ago next
I've been wanting to try out Tensorflow.js for a while now. I'm inspired to finally give it a shot after seeing this app.
username1 4 months ago prev next
Have you considered adding a feature where the app can guess your gender as well? That would definitely make it more useful for some applications.
username8 4 months ago next
I think adding a gender guessing feature would definitely make the app more interesting and useful. I'll definitely consider adding it in the future.
username8 4 months ago next
I'm glad you like the idea of adding a gender guessing feature. I'll definitely keep it in mind as I continue to develop the app.
username3 4 months ago prev next
How did you preprocess the audio files for training? I've always found that step to be the most challenging.
username4 4 months ago next
I used a simple feature extraction library to extract relevant features from the audio files, and then used a sliding window to generate more data. It's definitely not perfect, but it worked well enough for my purposes.
username5 4 months ago prev next
I built a similar app using C++ and OpenFrameworks a few years ago. It was a ton of work, but really rewarding in the end. I'm glad to see people building similar apps with more modern tools.
username10 4 months ago next
I'd love to see a side-by-side comparison of your C++ app and this Tensorflow.js app. It would be really interesting to see how the two technologies compare in terms of performance and ease of use.
username10 4 months ago next
I'll work on putting together a comparison in the near future. I'm curious to see the results myself.
username7 4 months ago prev next
How did you handle the model evaluation and selection process? Did you try out multiple architectures and compare their performance?
username9 4 months ago next
I tried out a few different architectures and ended up going with a simple CNN. I found that it performed well enough for this kind of app, and didn't require a lot of computational resources.
username11 4 months ago prev next
How did you ensure that the app could handle different accents and vocal styles? That has always been a challenge for me when building similar apps.
username12 4 months ago next
I made sure to include a diverse set of voices in the training data, which seemed to help the app handle different accents and vocal styles. It's definitely something that requires careful attention when building this kind of app.