Ariosa, our backbone engine for analog modeling, combines several advanced concepts in the field of modern artificial intelligence (AI), including differentiable digital signal processing (DDSP), machine learning (ML) and deep neural networks (DL). Leveraging the capability of AI, Ariosa not only precisely emulates the target analog machine without explicitly knowing the circuit schema, but also captures its characteristics in a delicate way.
Introducing the British Kolorizer
The British Kolorizer (BK), the first plugin made by Ariosa, is designed to digitize and optimize an analog British tube amplifier. Instead of analyzing the circuit, Ariosa was trained to fit the sound output of analog hardware directly. When modeling an amplifier, one essential feature is the degree of coloring responding to the input signal level, which makes the sound more diverse in timbre. To achieve this, the dynamics of the input signal for training is carefully designed.
The Sound of British Kolorizer
Both Fat and Brightness knobs are designed to color the sound but targeting different frequency bands. Bringing up the Fat knob boosts lower frequencies, which brings warmth and richness into the music; when raising the Brightness knob value, the higher frequencies will be further colored, which conveys a feeling of lightness. The combination of the two knobs are fully sampled, which enables BK to mimic the behavior of British tube amplifier more accurately.
BK was validated on an unseen audio file with 96k Hz sampling rate. Figure 1 shows that the signal rendered from BK plugin (Prediction) is nearly identical to the original British tube amplifier (Target). When comparing the two in a finer scale (shown in Figure 2), the result is further justified. In Figure 3, it can be seen that the frequency response of BK is very close to the target. In terms of metric, the MAE (mean absolute error) loss is around 0.1%, which indicates the ability of BK to faithfully digitalize the original hardware.
There are several disadvantages in the existing modeling methods: Black-box models such as Hammerstein-Wiener models, Volterra series and pure deep neural networks all have intensive computational overhead, so trade-off in quality usually needs to be considered; white-box models or traditional signal processing methods might lower the computational complexity through strong assumptions and prior knowledge, but always fall short of dynamics.
Compared to current works, Ariosa inspired plugins are well-balanced in terms of model size, computational performance, latency and quality. The robust performance is accomplished through the grey-box-like nature of Ariosa and the in-house math library based on C++ developed by the engineering team. Although the internal working sampling rate is 96K Hz, CPU and RAM consumption are still acceptable, and GPUs are apparently unnecessary.
Another spotlight of the Ariosa engine is its ability to automatically identify the latency, which could be introduced by the recording process or the analog machine itself. In the case of the British Kolorizer, the measured latency is approximately 300 samples under 96K Hz sampling rate. By identifying and reducing it, we can make those analog machines with ultra long latency become possible in real-time applications.
Ariosa is a new analog modeling technology based on AI. This core engine has high fidelity and preserves characteristics of the hardware. We also optimize the computational performance and reduce the latency for the plugin to be available on a regular laptop. British Kolorizer is the first product modeled by Ariosa. The quantitative evaluation demonstrates the success of emulating the tube amplifiers faithfully.
We are expanding Ariosa on analog machines of various audio effects and expecting to make high end music production more accessible for the public.
Posted by Master Tones Research Team