Master Thesis – Core Tech Media
Axis Communications AB · Sweden, Lund
Job TitleMaster Thesis – Core Tech Media
Job DescriptionWe have four ideas for you to pick from:
1) Knowledge distillation for audio classification task
2) Evaluating Generated and Simulated Training Data for Sound Event Classification
3) Energy-Efficient Audio Classification for (Battery-Powered) Edge Devices
4) Lightweight Speaker Anonymization for Intercoms
Category
Computer Science, Applied Mathematics, Technical Physics, Electronics Engineering or similar
Scope
Two Master students, 30hp each
Idea 1 - Knowledge distillation for audio classification taskKnowledge distillation is a machine learning technique that transfers knowledge from a large, complex model (the teacher) to a smaller, more compact model (the student).
Our Audio Classification model runs on our camera products (“edge hardware”). Therefore it is a smaller model, many order of magnitudes smaller than state-of-the-art classification models capable of running on big GPUs.
Relevant courses
C/Python, Machine Learning/AI
Goals
Explore using larger models on PC to improve a smaller model, to boost performance above what should be possible with the on edge model sizes.
Idea 2 - Evaluating Generated and Simulated Training Data for Sound Event ClassificationAudio classification can detect events like glass breaking, gunshots, screams, or alarms. These events share multiple problems though: they are rare, hard to record, and often privacy-sensitive.
Synthetic data offers a way around this bottleneck. Three families of methods have matured:
- Procedural/physical simulation: Clean event recordings are combined with simulated room acoustics, for example room impulse responses from software tools like pyroomacoustics. Microphone characteristics and background noise are added on top.
- Soundscape synthesis: Libraries such as Scaper mix foreground events into background scenes with controlled timing, SNR, and overlap.
- Generative models: Text-to-audio models such as AudioLDM, AudioGen, and Stable Audio generate novel event samples from prompts.
However, the extent to which each approach generalizes to real-world conditions, the proportion of real data it can substitute, and the circumstances under which synthetic data is detrimental remain largely unexplored.
Relevant courses
C/Python, Machine Learning/AI
Goals
Research Questions:
- How does classifier performance on real-world test data change when real training data is partially or fully replaced by synthetic data?
- Which synthesis strategy (physical simulation, soundscape mixing, or generative models) gives the best real-world generalization per unit of effort?
- Can a small amount of real data combined with synthetic pre-training match a model trained on a much larger real dataset?
- Can we predict the domain gap in advance, for example with Fréchet Audio Distance (FAD), and does FAD correlate with downstream accuracy?
Expected Contributions:
- A systematic, reproducible comparison of synthetic audio strategies for rare-event classification.
- Practical guidelines on how much real data is still needed and which synthesis method to use.
- A pipeline for generating and evaluating synthetic audio datasets.
Idea 3 - Energy-Efficient Audio Classification for (Battery-Powered) Edge DevicesAudio classification can detect glass breaking, gunshots, aggression, or machinery faults, which makes it a strong complement to video in security and IoT systems. Running these models continuously on battery-powered devices is impractical because always-on listening drains energy quickly. Most research optimizes for accuracy or latency. Far less addresses the full energy budget of the sensing pipeline, from microphone to decision. This thesis focuses on that gap.
How can a battery-powered device run continuous audio classification while minimizing average power consumption and keeping detection performance acceptable?
The average power of an always-on system can be modeled as: Here, is the duty cycle, and is the active power of each pipeline stage (sensing, feature extraction, inference, wake-up, communication). Lowering means reducing either how often each stage runs or how much power it draws while running.
Goals
Research questions:
- Cascaded detection: How much energy does a multi-stage pipeline save? Such a pipeline would use an ultra-low-power stage, such as energy thresholding or a tiny neural network, to gate a larger classifier.
- Model compression: How does quantization (int8 or lower), pruning, and knowledge distillation affect the trade-off between accuracy and energy per inference?
- Front-end optimization: Can cheaper features or lower sample rates and frame rates cut computation without a significant accuracy loss? Examples include fewer mel bands, learned filterbanks, or raw-waveform models.
- Adaptive operation: Can the system change its duty cycle or model size based on context, such as time of day, background noise, or battery level?
Expected contributions:
- A benchmark that measures accuracy against real energy use for compressed audio models on embedded hardware.
- Design guidelines for cascaded, always-on audio pipelines, quantifying where energy is actually spent.
- A prototype adaptive system showing longer battery life with minimal loss in detection quality.
Idea 4 - Lightweight Speaker Anonymization for IntercomsNetwork intercoms and emergency help points are increasingly connected to recording, cloud services, and remote operators. Recordings help with evidence, quality assurance, and incident review, but a voice is biometric data. Under GDPR, voice recordings that can identify a person are personal data, and voiceprints used for identification fall into special categories. This creates tension: organizations want to keep the conversation but not necessarily the identity of everyone involved.
Speaker anonymization addresses this by changing who appears to be speaking while preserving what is said. The VoicePrivacy Challenge (2020, 2022, 2024) has established a shared framework for the field, with baselines ranging from signal processing methods such as McAdams coefficient–based formant shifting to neural pipelines that extract linguistic content, replace the speaker embedding, and re-synthesize speech.
The strongest of these methods were designed for offline processing on GPUs, so they don't suit intercoms.
An intercom must:
- Run on embedded hardware with tight memory and compute budgets
- Operate in real time and full duplex, where ITU-T G.114 recommends keeping one-way delay below about 150 ms for natural conversation
- Handle narrowband or wideband audio, codec artifacts, outdoor noise, and reverberant entrances
What the field still lacks is a systematic study of how much anonymization strength can be retained once these methods are compressed to fit within the latency and compute budget of an embedded device.
Relevant courses
C/Python, Signal Processing, Machine Learning/AI
Goals
Research Questions:
- Privacy–utility trade-off: How do lightweight anonymization methods compare with state-of-the-art VoicePrivacy baselines in identity protection and speech intelligibility?
- Edge feasibility: Which methods meet real-time constraints (latency, memory, and compute) on representative embedded hardware, and what does it cost in privacy or quality?
- Attacker robustness: How well do lightweight methods hold up against an informed attacker who knows the anonymization method and adapts their speaker verification system to it?
- Domain robustness (optional): How do intercom conditions, such as narrowband audio, codecs, background noise, and reverberation, affect both anonymization strength and intelligibility?
OK, I am interested! What do I do now?
You are valuable to us – how nice that you are interested in one of our proposals! There are a few things for you to keep in mind when applying.
- Applications are accepted in both Swedish and English, and you apply via the proposal advert.
-