WAVDSP Neural Labs has introduced Primary Vx, a new real-time, low-latency AI vocal cleanup plugin. This innovative software is designed to detect and separate human voices from various background noises and stage bleed, ensuring clear speech and singing in demanding acoustic environments. Primary Vx aims to provide a pristine vocal signal, which is crucial for both live performance and studio recording scenarios. Furthermore, its advanced capabilities extend to feedback protection, making it a comprehensive tool for audio engineers. This plugin represents a significant advancement in vocal processing technology.

Core AI Pipeline and Three-Stage Cleanup

Unlike traditional processing methods that rely on IIR or FIR EQ, dynamic EQ, or phase tuning, Primary Vx utilizes a custom Pure AI pipeline. This unique approach allows for voice-aware cleanup through a focused three-stage chain. First, Vx Source Detection identifies the vocal source in real-time, enabling the engine to follow the performance accurately rather than the surrounding bleed. Subsequently, Source Separation works to eliminate unwanted background instruments, room ambience, and stage bleed from the vocal path. Finally, the Feedback Protection stage helps engineers push more gain before feedback occurs, significantly lowering the risk of ringing from the PA system. Consequently, this multi-stage processing ensures a cleaner and more stable vocal signal.

Addressing Real-World Challenges

Primary Vx is specifically engineered to tackle common challenges faced by vocal microphones in real-world settings. For instance, it effectively removes transient spill from drums, harsh high-end bleed from cymbals, midrange wash from guitar amps, and low-end rumble from bass amps. Moreover, it addresses room reflections, reverb tail buildup, and PA system bleed, all of which typically contaminate vocal signals. By pushing this noise away, the plugin ensures the voice remains the central focus. This results in cleaner live vocals and production-ready captures, reducing distracting bleed at the source. Therefore, it is an invaluable tool for challenging recording and live environments.

Advanced Training and Optimization

The development of Primary Vx involved extensive and specialized AI training. The Neural Labs team created proprietary datasets, specifically recorded for vocal separation and feedback protection, as off-the-shelf public datasets proved unsuitable. This rigorous process included curating 2.5 TB of licensed studio stems, multitrack rehearsals, and field captures. After meticulous review and trimming, an active training pool of 1.9 TB was used. The training itself consumed approximately 1,680 GPU-hours for the main vocal separation network and an additional 420 GPU-hours for fine-tuning the voice-activity detection and feedback-aware guard modules. This extensive effort ensures the model performs reliably even when the vocal is buried under significant stage noise.

System Compatibility and Performance

Primary Vx is currently available as an audio plugin for macOS 11.0 or later, supporting VST3, AU, and AAX formats. It is optimized for both Apple Silicon and Intel Macs, with Apple Silicon Macs particularly benefiting from Core ML and Apple Neural Engine acceleration for enhanced performance. While Windows support is not yet available, it is planned for a future release. The plugin boasts low internal latency, under 4.7 ms, making it suitable for Front of House (FOH) workflows and other live applications where quick response times are critical. Users can run multiple instances, depending on their CPU capabilities, and recommended host software includes SuperRack Performer and LiveProfessor. Additionally, online activation is required once, allowing offline use thereafter.

Features

• Vx Source Detection tracks vocal source in real time, not surrounding bleed.
• Source Separation removes instruments, ambience, and stage bleed from the vocal path.
• Feedback Protection for increased gain before feedback without vocal quality loss.
• Low-latency operation (under 4.7 ms internal latency).
• Custom Pure AI pipeline, no traditional EQ or filters.
• Optimized for Apple Silicon with Core ML and Apple Neural Engine acceleration.
• Supports VST3, AU, and AAX formats on macOS.

Price
License: $699

More info: WAVDSP | Primary Vx