http://mohammedkhalilia.com/tags/speech-and-signal-processing/Speech and Signal ProcessingHugoBlox Kit (https://hugoblox.com)en-usMon, 11 Jan 2021 00:00:00 +0000http://mohammedkhalilia.com/media/icon_hu_e7e672982174d01f.pnghttp://mohammedkhalilia.com/tags/speech-and-signal-processing/http://mohammedkhalilia.com/projects/voice-style-transfer/Mon, 11 Jan 2021 00:00:00 +0000http://mohammedkhalilia.com/projects/voice-style-transfer/<p>Voice style transfer, also known as voice cloning, create the same voice characteristics as of the original speaker while uttering different words in different languages. To do that, a variational auto-encoder model was trained on two datasets:</p> <h5 id="voice-conversion-toolkit-vctk">Voice Conversion Toolkit (VCTK)</h5> <ul> <li>44 hours</li> <li>109 speakers</li> </ul> <h5 id="voxceleb1-and-voxceleb2">VoxCeleb1 and VoxCeleb2</h5> <ul> <li>2,000+ hours</li> <li>7,000 speakers</li> </ul> <p>The above data, which contains waveforms is converted to mel spectrograms using Short-Time Fourier Transform. <figure > <div class="flex justify-center "> <div class="w-full" > <img alt="alt" srcset="http://mohammedkhalilia.com/projects/voice-style-transfer/fft_hu_9f1cb69ded5d4f92.webp 320w, http://mohammedkhalilia.com/projects/voice-style-transfer/fft_hu_764a88c2c35cdc64.webp 480w, http://mohammedkhalilia.com/projects/voice-style-transfer/fft_hu_80a6dc05ae176835.webp 760w" sizes="(max-width: 480px) 100vw, (max-width: 768px) 90vw, (max-width: 1024px) 80vw, 760px" src="http://mohammedkhalilia.com/projects/voice-style-transfer/fft_hu_9f1cb69ded5d4f92.webp" width="760" height="77" loading="lazy" data-zoomable /></div> </div></figure> </p> <p>Then a variational auto-encoder called AutoVC, is trained on the data.</p>