Independent Research
I conduct independent applied research at the intersection of audio engineering, machine learning, and human–AI interaction. My current work focuses on how AI-assisted and agentic systems can support professional audio workflows while preserving intelligibility, speaker identity, editorial control, reproducibility, and human judgment.
Evaluation of machine-learning systems for dialogue restoration, noise reduction, source separation, speech enhancement, and perceptual audio quality.
The goal is not only to improve objective quality metrics, but also to understand how automated processing affects intelligibility, naturalness, speaker characteristics, and real-world production decisions.
Research into speech-to-text, speaker diarization, transcript-based media editing, semantic search, and the relationship between linguistic representations and non-destructive audio/video editing workflows.
A central question is how accurately language-level decisions can be translated into reliable timeline-level edits.
Exploration of how AI systems can assist audio engineers, editors, musicians, podcasters, and other creative professionals without removing meaningful human control.
This includes human-in-the-loop workflows, explainability, confidence-aware automation, reversible editing, and practical evaluation of AI-assisted production systems.
Investigation of agentic AI systems for experiment orchestration, literature analysis, benchmarking, software development, documentation, and reproducible computational research.
The focus is on workflows in which AI agents can assist with complex research tasks while keeping methods, decisions, data, and outputs traceable and auditable.
To what extent can locally executable AI systems improve spoken-word production workflows while maintaining speech intelligibility, perceptual quality, speaker identity, editorial accuracy, and meaningful human control?
Current speech and media AI systems often perform individual tasks—such as transcription, noise reduction, speaker detection, source separation, or semantic analysis—very well in isolation.
However, fewer studies examine how these systems behave when combined into complete production workflows.
This research investigates the performance, limitations, and practical trade-offs of integrated AI-assisted speech production pipelines.
The research uses a computational and experimental benchmarking approach.
Multiple models and processing pipelines will be evaluated using controlled spoken-word recordings as well as representative real-world production material.
Where appropriate, evaluation will combine objective measurements with structured human listening and editorial assessments.
Special attention will be given to cases where objective quality metrics and human perception disagree.
All experiments will aim to be reproducible, with model versions, processing parameters, evaluation methodology, and computational environment documented whenever licensing and data constraints permit.
Establish a reproducible evaluation framework for speech enhancement, transcription, diarization, and transcript-driven editing systems.
Compare individual models and integrated pipelines across different recording conditions and production scenarios.
Investigate agent-assisted workflows that can execute experiments, compare results, identify failure cases, and generate reproducible research artifacts while retaining human oversight.
Consolidate findings into openly documented reference workflows, benchmarking tools, research notes, and technical reports.
Software for comparing speech-processing and AI-assisted editing workflows.
Reproducible procedures for testing speech restoration, transcription, diarization, and editing systems.
Documented AI-assisted production pipelines that can be reproduced by other researchers and practitioners.
Redistributable test material and annotations where data rights permit publication.
Public documentation of methodology, experiments, results, limitations, and negative findings.
Tools developed during the research that may be useful to the wider audio, speech, and AI research communities.
Research results should be repeatable wherever technically and legally possible. Model versions, parameters, datasets, and evaluation methods should therefore be documented.
Numerical metrics alone are not sufficient for evaluating creative and perceptual audio systems. Human listening and editorial assessment remain essential components of the research.
AI should support informed creative and technical decisions rather than obscure them. Systems should favor reversible actions, transparent decision-making, and meaningful user control.
Local processing and privacy-preserving workflows are particularly important when working with speech recordings, unpublished media, and professional production material.
Whenever intellectual-property, licensing, and privacy constraints allow, software, evaluation methodology, findings, and research artifacts will be made publicly accessible.
Sercan Solmaz
Sercan Solmaz is an audio engineer, music producer, educator, musician, and developer working across professional audio, music technology, media production, and artificial intelligence.
His professional background spans approximately two decades of work with recording systems, music production, audio technology, post-production, education, and creative technology.
His current technical work focuses on AI-assisted audio and video systems, local machine-learning workflows, agentic software, speech technologies, and tools for creative professionals.
This combination of long-term professional audio practice and software experimentation informs a research approach centered on real production environments rather than isolated model performance.
I am interested in collaboration with researchers, engineers, and organizations working in areas including:
I am also interested in participating in benchmark design, early-stage technology evaluation, research interviews, experimental tool testing, and open research initiatives.
Sercan Solmaz
Independent Research — Audio AI & Human–AI Creative Systems