Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Bachelor/Master Thesis: Alcohol intoxication detection based on speech
Traffic accidents cause over a million deaths every year, of which a large fraction is attributed to drunk driving. We believe that voice-based detection systems are the natural way of inferring whether a driver is drunk and to reduce traffic accidents.
Keywords: AI
machine learning
audio processing
driver state
signal processing
speech processing
The fact that an increasing number of functions in the automobile is and will be controlled by the speech of the driver rises the question of whether this speech input may be used to detect a possible alcoholic intoxication of the driver. For that matter, we have access to a large dataset of sober and intoxicated speakers in the car, the Alcohol Language Corpus (ALC) [1]. This dataset has the following characteristics:
- 162 participants (male and female) who were recorded sober and under the influence of alcohol
- All recordings available
- All recordings were completed in the car environment with professional equipment
- Speech recordings for each participant: Sober (30 min. with 60 speech tasks) and drunk (15 min. with 30 speech tasks)
- Speech tasks: Dialogues, commands, reading, free speaking,... all with varying length
- The content is available both as audio and transcribed
[1]: Schiel, Florian, Christian Heinrich, and Sabine Barfüsser. "Alcohol language corpus: the first public corpus of alcoholized German speech." Language resources and evaluation 46.3 (2012): 503-521.
The fact that an increasing number of functions in the automobile is and will be controlled by the speech of the driver rises the question of whether this speech input may be used to detect a possible alcoholic intoxication of the driver. For that matter, we have access to a large dataset of sober and intoxicated speakers in the car, the Alcohol Language Corpus (ALC) [1]. This dataset has the following characteristics:
- 162 participants (male and female) who were recorded sober and under the influence of alcohol - All recordings available - All recordings were completed in the car environment with professional equipment - Speech recordings for each participant: Sober (30 min. with 60 speech tasks) and drunk (15 min. with 30 speech tasks) - Speech tasks: Dialogues, commands, reading, free speaking,... all with varying length - The content is available both as audio and transcribed
[1]: Schiel, Florian, Christian Heinrich, and Sabine Barfüsser. "Alcohol language corpus: the first public corpus of alcoholized German speech." Language resources and evaluation 46.3 (2012): 503-521.
Audio processing and selection of relevant features.
Use state-of-the-art machine learning methods to infer the intoxication level of people based on the derived speech features.
You would implement different machine learning approaches and evaluate them against each other. The machine learning models can range from more conventional pipelines with the need to create granular speech-related features (e.g. vocal acoustics, prosodic and source features, fundamental frequency, first and second format frequency, or mel-frequency cepstrum (MFC)) over end-to-end prediction models to re-use existing speech networks and use them via transfer learning.
Audio processing and selection of relevant features.
Use state-of-the-art machine learning methods to infer the intoxication level of people based on the derived speech features.
You would implement different machine learning approaches and evaluate them against each other. The machine learning models can range from more conventional pipelines with the need to create granular speech-related features (e.g. vocal acoustics, prosodic and source features, fundamental frequency, first and second format frequency, or mel-frequency cepstrum (MFC)) over end-to-end prediction models to re-use existing speech networks and use them via transfer learning.
Kevin Koch (kevin.koch@unisg.ch) or Gisbert Teepe (gteepe@ethz.ch)
The thesis is a joint project between the Bosch IoT Lab (A Cooperation of ETH Zürich, University of St. Gallen and Bosch, https://www.iot-lab.ch/) and the Center for Digital Health Interventions (a joint initiative of ETH Zürich and University of St. Gallen, https://www.c4dhi.org/).
Please contact one of us with your CV, a short statement of motivation, and your current transcripts of records (bachelor & master).
Kevin Koch (kevin.koch@unisg.ch) or Gisbert Teepe (gteepe@ethz.ch)
The thesis is a joint project between the Bosch IoT Lab (A Cooperation of ETH Zürich, University of St. Gallen and Bosch, https://www.iot-lab.ch/) and the Center for Digital Health Interventions (a joint initiative of ETH Zürich and University of St. Gallen, https://www.c4dhi.org/).
Please contact one of us with your CV, a short statement of motivation, and your current transcripts of records (bachelor & master).