#seedvc

We've released Voice Canvas, a tool that lets you design voices using 8 sliders, under the MIT license. We'll guide you through its capabilities, development insights, and how to set it up on Mac, Linux, and Windows, all the way to comparing your first created voices.

A method to visualize a set of voice anchors for voice conversion by showing their coverage across semantic axes and a PCA scatter plot of speaker embeddings. This approach highlights gaps and biases in a voice library, helping determine which voices to add next.

Upsampling 22.05kHz audio to 44.1kHz does not generate high-frequency content. Switching to an F0-conditioned 44k model results in approximately 7.4% energy in the >11kHz range, as measured in practice — we explain the decisive difference from mere resampling.

Labeling voice anchors without metadata by matching them with known speakers based on speaker embedding similarity, estimating the origin based on voice quality alone, and turning unknown materials into assets.

Seed-VC's voice conversion system follows a 4-stage pipeline: whisper (meaning) + campplus (speaker identity) + CFM/DiT (diffusion-based mel-spectrogram generation) + BigVGAN (vocoder). This design separates speaker identity, content, and prosody into distinct modules, and we'll walk through the implementation step by step.

Because diffusion models start from random noise, results can vary slightly even with the same input. This article explains how to achieve deterministic generation by fixing the seed with torch.manual_seed, and discusses its practical importance for A/B testing, parameter sensitivity analysis, and ensuring reproducibility.

A record of attempting to implement a slider that adds breath to voice, trying various methods such as addition, cross-synthesis, and model-generated non-periodic components, before ultimately giving up and removing it from the UI, as breath leaks and breath noise proved to be physically inseparable from noise.

A discussion on selecting voice conversion anchors (speaker materials) using objective metrics rather than just listening. I use a three-stage filter—mixed-speaker detection, prosody-based naturalness, and SQUIM/PESQ quality gates—to improve minimum PESQ scores from 1.24 to 2.31.

Only the voice-converted audio was inexplicably stretched during playback—the cause was not the model itself, but Whisper's 30-second limit used for semantic extraction. We present the investigation from symptoms to root cause and the solution through chunking.