Aditya Tb
Machine Learning Research Engineer | Music, Audio, Speech Tech
ML Research Engineer in Speech/Audio/Music with a multi-discplinary background in Audio Engineering and Music Technology. Expertise in Signal Processing, Generative AI and Deepfake Detection for Audio. Currently researching countermeasure systems to assist in ethical use of generative AI systems. Currently based in Canada.
- Location
- Ottawa, Ontario, Canada
- aditya.tirumala@mail.mcgill.ca
- Website
- https://adityatb.github.io/
- GitHub
- adityatb
- in/adityatb
Experience
– present
Machine Learning Research Engineer at Resemble AI
Currently leading the research and development of the deepfake detection system for audio.
Highlights
- State of the art accuracy on real‑world classification of speech+audio. EER < 6% on Franhoufer's In the Wild dataset.
- Improving and maintaining changes to production via CI/CD and replicate AI.
- Deep Learning R&D: Neural Upsampling, Audio Watermarking, Audio Deepfake Detection
–
Founder, Sound Designer at Sample Culture
Location sound recording, sound design and audio production house catering to the needs of advertising agencies, independent filmmakers and musicians in India.
Highlights
- Constructed a soundproof recording studio for Sample Culture, with a flat response b/w 20Hz-20kHz A-weighted.
- Award winning compositions, location sound and sound design for independent films and advertising agencies.
Education
–
Masters in Music Technology from McGill University
Courses
- Digital Sound Synthesis and Audio Processing
- Computational Modeling of Musical Acoustic Systems
- Music Information Acquisition, Preservation, and Retrieval
- Input Devices for Musical Expression
- Gestural Control of Sound Synthesis
–
Bachelor in Audio Engineering from School of Audio Engineering
Courses
- Procedural Audio Design with Max/MSP, Pure Data
- Thesis: Procedural Footstep synthesis for video games
–
Bachelor in Electronics and Communication Engineering from Jawaharlal Nehru Technological University (JNTU)
Publications
Defining a vibrotactile toolkit for digital musical instruments: characterizing voice coil actuators, effects of loading, and equalization of the frequency response in Journal on Multimodal User Interfaces 14, Springer
Given the extreme requirements of musical performances, there is a need for solutions allowing for independent control of frequency and amplitude over a wide frequency bandwidth (40–1000 Hz) and low harmonic distortion, so that flexible and high-quality vibrotactile feedback can be displayed. In this paper, we evaluate cost-effective and portable solutions that meet these requirements.
Autoregressive parameter estimation for equalizing vibrotactile systems in International Workshop on Haptic and Audio Interaction Design-HAID2019
For advanced musical use, it is essential to perform the characterization of the amplifiers and actuators involved, as well as the equalization of their overall frequency response characteristics, a step typically implemented with the help of manually configured parametric equalizers. This paper proposes an autoregressive method that automatically estimates minimum-phase filter parameters, which by design, remain stable upon inversion.
Skills
- Machine Learning
- Level: MasterKeywords:
- Audio Signal Processing
- Level: MasterKeywords:
- Backend Development
- Level: IntermediateKeywords:
Interests
- Real-Time Differential Signal Processing
- Keywords:
- Audio Deepfake Detection
- Keywords:
- Generative Music and SFX
- Keywords: