Who Do I Sound like? Showcasing Speaker Recognition Technology by Youtube Voice Search

Ville Vestman, Bilal Soomro, Anssi Kanervisto, Ville Hautamäki, Tomi Kinnunen

Abstract

Bridging the gap between complex academic research and public understanding is a significant challenge in the field of speaker recognition. This paper introduces a web-based "celebrity voice match" application designed to make speaker identification technology accessible and engaging. Leveraging the extensive VoxCeleb dataset and a highly optimized i-vector extraction backend, the system identifies a user's closest celebrity vocal match with 93% accuracy in under a second. We discuss the system architecture, the importance of low-latency processing for user experience, and the results of public testing.

Introduction

Traditional speaker recognition demonstrations often suffer from high user friction, requiring tedious enrollment phases, and provide abstract numerical feedback that fails to engage non-technical audiences. To address these issues, we developed a "voice search" application that compares a user's voice to thousands of pre-enrolled celebrities from YouTube. This approach provides immediate, visual, and relatable feedback, turning a technical verification task into an entertaining social experience.

The 'Who Do I Sound Like?' web interface, showing the recording tool and celebrity match results.

System Architecture and Methods

The application utilizes a closed-set speaker identification architecture optimized for speed and robustness.

Feature Extraction and Backend

The system extracts 20 Mel-frequency cepstral coefficients (MFCCs) along with delta and double-delta coefficients. An energy-based detector is employed to discard silent frames. To ensure sub-second response times on standard hardware, we implemented an optimized i-vector backend. Specifically, we used Probabilistic Principal Component Analysis (PPCA) for i-vector extraction, allowing for high-dimensional data compression into an 800-dimensional vector via a precomputed projection matrix. This method bypasses the computationally expensive matrix inversions typical of traditional i-vector extraction.

Scoring and Web Integration

Scoring is performed using a simplified Gaussian Probabilistic Linear Discriminant Analysis (G-PLDA) model. The user's i-vector is scored against over 900,000 pre-calculated celebrity i-vectors. The front end, built with JavaScript, handles in-browser 16kHz audio recording, while the Python-based backend manages high-performance linear algebra using NumPy and SciPy.

Datasets: The Role of VoxCeleb

The system's effectiveness relies on the VoxCeleb1 and VoxCeleb2 corpora, which contain over a million speech clips from more than 7,000 celebrities. Beyond training data, VoxCeleb provides YouTube metadata (IDs and timestamps). This allows the application to embed the specific YouTube video of the matched celebrity, enabling users to immediately compare their voice to the result.

Experimental Results

The system was evaluated for accuracy, latency, and robustness.

Identification Accuracy and Utterance Length

In controlled tests, the system achieved a top-1 accuracy of 93% and a top-5 accuracy of 100%. A critical finding was the relationship between utterance length and performance. While 5 seconds of speech provides over 70% accuracy, approximately 9 seconds of audio is necessary to reach near-optimal performance levels.

Identification accuracy as a function of test utterance length for top-1, top-3, and top-5 positions.

Computational Efficiency and Robustness

The backend processing time averages 661ms, leading to a total user-perceived response time of approximately 2.5 seconds (including network overhead). The system also demonstrated robustness against "replay channel" distortion; when recording YouTube clips played through a Bluetooth speaker via a smartphone, top-1 accuracy remained high at 87%.

Public Reception and Limitations

During "The European Researchers’ Night," the demo was tested by approximately 150 participants with overwhelmingly positive feedback. However, several limitations were identified:

  • Demographic Bias: The VoxCeleb dataset primarily features adults, leading to poor matching for children.
  • Browser Compatibility: Variations in MediaRecorder API support limited the demo's functionality to an estimated 50-70% of user devices.

Conclusion

The "Who Do I Sound Like?" application successfully demonstrates that speaker recognition technology can be made engaging for the general public through intuitive UI design and celebrity-based matching. Future work will focus on expanding the search criteria to include age, emotion, and language, as well as improving cross-browser compatibility.

The source code for this platform is available on GitHub to support further public-facing research demonstrations.

Summary updated

Abstract

The popularization of science can often be disregarded by scientists as it may be challenging to put highly sophisticated research into words that general public can understand. This work aims to help presenting speaker recognition research to public by proposing a publicly appealing concept for showcasing recognition systems. We leverage data from YouTube and use it in a large-scale voice search web application that finds the celebrity voices that best match to the user's voice. The concept was tested in a public event as well as "in the wild" and the received feedback was mostly positive. The i-vector based speaker identification back end was found to be fast (665 ms per request) and had a high identification accuracy (93%) for the YouTube target speakers. To help other researchers to develop the idea further, we share the source codes of the web platform used for the demo at https://github.com/bilalsoomro/speech-demo-platform.
BibTeX
@inproceedings{vestman2018who,
  author = {Ville Vestman and Bilal Soomro and Anssi Kanervisto and Ville Hautamäki and Tomi Kinnunen},
  title = {{Who Do I Sound like? Showcasing Speaker Recognition Technology by Youtube Voice Search}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2018},
  doi = {10.1109/icassp.2019.8683272},
}
Figure