Reproduce “DarkStream: real-time speech anonymization with low latency” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Waris Quamer, Ricardo Gutierrez-Osuna

DarkStream is a real-time, low-latency speech anonymization system that conceals speaker identity while preserving speech intelligibility and emotional content. It achieves an end-to-end latency of 203ms with 140ms lookahead and demonstrates a semi-informed EER of 22.68% when using k-means quantization, performing comparably to some offline baselines.

Key insights

  • Integrating a limited lookahead buffer and causal transformer-based contextual layers significantly improves content encoding accuracy and speech utility (WER, UAR) even under strict latency constraints.
  • K-means quantization of content embeddings is highly effective in boosting speaker privacy (increasing EER to near-chance levels in some scenarios), albeit at a measurable cost to speech naturalness (lower MOS) and intelligibility (higher WER).
  • Direct waveform generation via a neural vocoder, combined with streaming causal architectures, enables end-to-end latency suitable for real-time interactive applications while maintaining high speech quality.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.