# CALM - Technical Paper **Everworks** **Neural feedback control, de-reverb and vocal protection** **Version | May 2026** --- ## 1. Executive Summary CALM is a professional real-time voice processor for live sound, broadcast, post-production and studio work. Its main goal is to reduce acoustic feedback, tonal resonances, room tail, background noise and external bleed without turning the voice into a dull or artificial signal. The philosophy behind CALM is simple: detect what threatens vocal stability or intelligibility, then act only where action is needed. To do that, CALM combines Mamba neural engines, multi-resolution spectral analysis, tonal detection, per-band Kalman memory and an integrated vocal DSP chain. CALM is not only a de-feedback tool, not only a de-reverb tool and not only a channel strip. It is a compact vocal chain designed to keep a voice useful, stable and present in real-world conditions. ### Product Keys | Area | Implementation | |---|---| | Product | CALM | | Brand | Everworks | | macOS formats | AU, VST3, Standalone | | macOS architecture | Universal Binary: Apple Silicon arm64 + Intel x86_64 | | Windows formats | VST3, Standalone | | Processing modes | Studio 1024, Balance 512, Live 256 | | Neural engines | CALM, SERENE, BLEED | | De-reverb engine | SERENE Mamba V4 Elite | | Anti-feedback engine | CALM neural + Tonal Lock + Kalman | | De-bleed engine | BLEED Mamba V3 | | Vocal DSP | Expanse, Resonance, DNS, Clean, Body, DeEss, Vocal Finisher, Glue | --- ## 2. Problem Statement In a real vocal chain, several problems often happen at the same time: - Acoustic feedback before the system reaches the required working level. - Narrow tones that appear when voice or music excites a critical frequency. - Room reverberation that reduces intelligibility. - External bleed from monitors, backline, audience or room ambience. - Constant background noise. - Sibilance, harshness or lack of body. - Low latency requirements in live situations without losing quality. Traditional processors usually attack only one of these problems. CALM combines several decision layers: - Neural model for broad spectral behavior. - Tonal detector for narrow and sustained resonances. - Tonal Lock for headset microphones, lecterns and DPA-style use cases. - Kalman Adaptive memory for confirmed risk bands. - Vocal DSP chain for final polishing. --- ## 3. General Architecture ```text Input | |-- Auto SRC when needed | |-- STFT analysis | | | |-- Studio 1024: maximum resolution | |-- Balance 512: low latency with CALM V8 | |-- Live 256: minimum latency | |-- Spectral feature extraction | | | |-- Bark / Mel features | |-- Tonality | |-- Peak stability | |-- Spectral flatness | |-- Decision engines | | | |-- CALM anti-feedback | |-- Tonal Lock | |-- SERENE de-reverb | |-- BLEED de-bleed | |-- Kalman Adaptive risk map | |-- Mask and gain fusion | |-- IFFT / Overlap Add | |-- Vocal Chain | |-- Expanse |-- Resonance / Soothe |-- Clean |-- DNS |-- Body |-- DeEss |-- Vocal Finisher |-- Glue |-- Auto Gain / Output Trim | Output ``` --- ## 4. Processing Modes ### Studio 1024 Maximum spectral resolution mode. It is best suited for studio, mixing, broadcast, post-production and situations where latency is not the main constraint. Recommended use: - Lead vocal in studio. - Podcast or recorded broadcast. - Cleaning already recorded material. - Systems where precision matters more than latency. ### Balance 512 Quality and latency compromise mode. This is the main path for the modern 512 route and uses the CALM V8 runtime with a 46-feature input. Recommended use: - Live sound with low-latency requirements. - Theatre, conferences and headset microphones. - Monitor systems with feedback risk. - Fast testing of the 512 model. ### Live 256 Minimum latency mode. It reduces spectral resolution but allows faster response in live situations. Recommended use: - In-ear monitoring. - Critical PA conditions. - CPU-limited systems. - Scenarios where latency is the first priority. --- ## 5. Neural Engines ### 5.1 CALM CALM is the main anti-feedback engine. It analyzes spectral energy, peak stability, tonality and the temporal behavior of the signal in order to create a selective reduction mask. The Balance 512 route uses CALM V8: - Input: 46 features. - Mask: 44 bands. - State: causal streaming. - Goal: act on feedback and resonances without destroying the voice. The Studio 1024 route keeps the most solid high-resolution response. The Live 256 route keeps the fast low-latency version. ### 5.2 SERENE Mamba V4 Elite SERENE is CALM's de-reverb engine. This version uses the Mamba V4 Elite model as the single user-facing SERENE route. Technical data: | Field | Value | |---|---| | Domain | 80 Mel | | Input | 82 features | | Mask | 80 bands | | Exported parameters | 284,880 floats | | Integrated model | serene_mamba_80mel_v4_best_elite.bin | V4 training validation snapshot: | Metric | Value | |---|---| | Tail reduction | -8.04 dB | | Direct loss | -1.33 dB | | Separation | 6.70 dB | | Elite score | 11.065 | Interpretation: SERENE is not designed to make the voice completely dry at any cost. Its goal is to reduce room tail while keeping the direct voice inside a musical range. ### 5.3 BLEED Mamba V3 BLEED is designed to reduce external spill and unwanted ambience around the voice. It is a separate engine from SERENE. Technical data: | Field | Value | |---|---| | Input | 46 features | | Mask | 44 bands | | Integrated model | calm_bleed_512_v3.bin | | Main use | De-bleed / spill suppression | BLEED does not replace an offline source separator. It is designed for real-time control, with gradual behavior and an independent amount control. --- ## 6. Tonal Lock Tonal Lock is the specialized layer for narrow and sustained tones. It is especially useful with: - DPA-style headset microphones. - Lavalier microphones. - Lecterns. - Theatre. - Conferences. - Voice with background music that excites fixed frequencies. How it works: 1. Searches for narrow local peaks. 2. Checks that they are stable and sufficiently tonal. 3. Allows part of the vocal protection to be bypassed when the peak is clearly a resonance. 4. Reinforces the CALM mask only in that area. 5. Feeds the Kalman Adaptive map when the detection is confirmed. Difference from regular CALM: - CALM protects the voice more and acts in a broader way. - Tonal Lock is more surgical with pure tones and narrow feedback. --- ## 7. Kalman Adaptive Kalman Adaptive is a per-band risk map. It is fed by real confirmed detections from CALM/Tonal Lock and by the internal engine state. The visual Kalman matrix is not decorative animation: cells light up with real per-band risk. If there is no risk, they should remain dark. Technical use: - Reduces thresholds in high-risk bands. - Reinforces suppression when a frequency comes back. - Helps CALM react faster to repeated resonances. --- ## 8. Vocal Chain The Vocal Chain can stay hidden or be expanded for advanced adjustments. It is designed as a finishing area, not as the center of the product. ### Expanse Expander/gate for reducing noise between phrases without aggressive chopping. Controls: - Threshold. - Depth. - Speed. ### Resonance / Soothe Dynamic resonance suppressor. It acts as a tonal smoothing layer that complements CALM. Controls: - Sensitivity. - Depth. - On/Off. ### DNS Noise reduction based on spectral estimation. Designed for air conditioning, constant noise, noisy preamps or background ambience. Controls: - Bias. - Attenuation. ### Clean Presence and clarity equalization with a musical approach. Controls: - Focus mode. - Focus gain. - Clarity mode. - Clarity gain. ### Body Low-body control for recovering weight without muddying the voice. Controls: - Frequency. - Amount. - Slope. ### DeEss Sibilance control with spectral mode and threshold, range, frequency and sharpness adjustments. ### Vocal Finisher Final vocal color: - Drive. - Opto. - Air. ### Glue Bus-style compression for gluing the final vocal signal. Controls: - Threshold. - Ratio. - Makeup. - Auto makeup. ### Auto Gain and Output Trim Auto Gain compensates level changes caused by processing. Output Trim allows final output adjustment for session integration. --- ## 9. Interface CALM is organized into three levels: 1. Top area: mode, Auto SRC, Ghost, Bypass and RTA. 2. Central area: CALM, Tonal Lock, SERENE, BLEED, sliders and main knob. 3. Vocal Chain: expandable advanced modules. The design goal is fast operation: - CALM in the center. - SERENE and BLEED with independent amounts. - Tonal Lock close to the protection area. - Secondary modules can be hidden to avoid visual overload. --- ## 10. Distribution ### macOS Universal Contents: - CALM Universal.component - CALM Universal.vst3 - CALM Universal.app Architectures: - Apple Silicon arm64. - Intel x86_64. ### Windows Contents: - CALM Windows.vst3 - CALM Windows.exe Architecture: - Windows x64. --- ## 11. Positioning CALM is designed for users who need real stability: - Live sound engineers. - Broadcast. - Professional podcast. - Theatre. - Conferences. - Voice studios. - Creators who need to clean voice quickly. It does not try to replace a full mixer or an offline restoration suite. Its strength is the real-time combination of anti-feedback protection, resonance control, moderate de-reverb, bleed control and vocal finishing. --- ## 12. Version Status CALM integrates: - CALM 512 V8. - SERENE Mamba V4 Elite. - BLEED Mamba V3. - Tonal Lock. - Real Kalman Adaptive. - Expandable Vocal Chain. - macOS Universal. - Windows VST3 + Standalone. This version is focused on what matters: a stable, clear and controlled voice, with a cleaner interface and fewer distractions.