Rahul Peter

• Working as a full-time intern at Trinnov Audio in Paris, supervised by Romain Couderc and Théophile Dupré, and co-supervised by Inria (Antoine Deleforge, Jean-Daniel Pascal, and Cedric Foy).
• Developing ML/DSP methods to extract loudspeaker (source) characteristics from room measurements, enabling improved calibration and equalisation.
• Focus on simulation, model training, and validation on real-world data within the research team headed by Miguel Moleron and Axel Coulon.

• Developed the Adversarial Masking (Adv-M) framework for robust sound source localization in the Spherical Harmonic (SH) domain, designing and implementing a GAN-based real-time binary mask generator to separate target speech from directional interference.
• Built and integrated the full ML pipeline: SH decomposition, adversarial mask generation, and downstream CNN-based DOA estimation; contributed to deriving and incorporating TDOA-informed model design choices.
• System demonstrated substantial gains (approximately 30% localization accuracy improvement and 44% RMSE reduction) and led to a paper accepted at the IEEE Asilomar Conference on Signals, Systems & Computers 2025.

• Backend Developer in the Online Payments Plural Team at PineLabs, collaborating with industry-leading merchants such as Flipkart, Amazon, and brands like Apple to build and integrate affordability and EMI features into the online payments platform, streamlining business processes.
• Enhanced Svix webhook service performance by implementing custom message serialization in Rust, reducing latency by 34%.
• Built parallel batch processing pipelines with exponential backoff for Dead Letter Queue handling.
• Developed retry logic with circuit breaker patterns in Kotlin and secure payload verification for financial events.
• Integrated metrics collection for delivery success rates and processing times.

• Session pianist, composer, and live FOH assistant.
• Produced, mixed, and mastered tracks; contributed to open-source audio tools.

Multimodal Lyric Intelligibility (ICASSP Cadenza)

Ranked 2nd out of 30 teams (intrusive track) and invited to present the paper at ICASSP 2026: NCC=0.69, RMSE=0.265. Compact hybrid system combining audio features with text-level metrics to assess lyric intelligibility in mixtures.

• Extracted ASR hypotheses (Whisper/Wav2Vec2/WavLM/Canary) and linguistic features (NLTK POS ratios, syllables/word), computed BERTScore against references, and fused scalar blocks with attention-pooled encoder embeddings.
• Implemented vocal separation (Demucs) and computed intrusive perceptual metrics (STOI, Zimtohrli) to model acoustic quality and lexical alignment.
• Emphasis on clean, reproducible pipelines for data processing, scoring, and model training.

Esperanto ASR Course Competition at Aalto

Ranked 3rd in class out of 15 teams. (WER=0.13, CER=0.04). Reproducible code: RP335/speech_rec_course_comp.

• Built NeMo manifests and fine-tuned a Conformer-Transducer with SpecAugment and partial layer freezing; averaged top-5 checkpoints for stability.
• Ensemble decoding with wav2vec2-large-xlsr-53-esperanto via NIST ROVER; applied rule-based phonetic post-processing (word-final devoicing) for consistent pronunciation.
• Modular scripts for normalization, inference, and submission; demonstrates multilingual text handling, hypothesis scoring, and reproducible research code.

Personalized Hearing Protection for First Responders

As Tech Lead for the Aalto PDP team, developed a smart audio module between radio/comms and headset to balance protection and intelligibility. The developed IP and algorithms are currently being integrated by Savox and Otos.

• Real-time DSP pipeline (WDRC, noise reduction, safety limiter, smart mixing) targeting <3ms latency on embedded hardware.
• Personalization using audiogram + PAR fit-check + noise dose tracking.
• Work is confidential; currently being implemented/patented by industrial partners.

DCASE 2025 SELD Challenge

Conformer ensemble for sound event localization and detection using synthetic data from SpatialScaper. Technical report

Query-by-Vocal-Imitation (QVIM) Challenge

Top-3 team finisher using MobileNetV3, PANNs, PaSST, and BEATs transformers on VimSketch. Focus on query-to-item matching, embeddings, and retrieval. qvim-aes.github.io | QVIM-Aalto

Voiced/Non-Voiced Detection in Speech

Research project under Prof. Anurag Nishad on detecting voiced vs. non-voiced segments using iterative Variational Mode Decomposition (VMD) to extract fundamental frequency components and their envelopes. Evaluated against Empirical Mode Decomposition and wavelet-based baselines on CMU Arctic and NOISEX-92 datasets under varied noise conditions to quantify robustness. Details