This website presents a collection of interactive demonstrations showcasing various speech technology applications. The demonstrations highlight the capabilities of modern speech models and provide examples of their practical use.

Speech input can be provided either through your microphone or by uploading an audio file. To use the microphone option, please grant your browser permission to access your microphone.

Once an audio sample has been provided, you can analyze either the entire recording or selected portions of it. Simply drag across the waveform to choose a specific segment for analysis.

The website is currently under active development, and you may occasionally encounter bugs or temporary service interruptions.

We invite you to explore the demonstrations and learn more about current speech technology capabilities.

We prioritize your privacy and take appropriate measures to protect your data. When you record or upload audio files, they are securely transmitted to our server for processing. Once processing is complete, the audio files are immediately and permanently deleted. We do not store personal audio data or associated information. Text inputs provided to the speech synthesizer may be temporarily stored on the server.

Last updated: July 2026

By accessing and using SpeechLab (the "Site"), you agree to the following terms:

1. Purpose of the Site

SpeechLab is a research demonstration platform providing interactive speech technology demos, including speaker recognition, speech analysis, speech synthesis, and fake speech detection. The Site is intended for educational and research purposes only.

2. User-Submitted Audio

• You may upload or record audio through the Site's demos.
• You are responsible for ensuring you have the right to use any audio you submit.
• Do not upload audio containing personal data of third parties without their consent.
• Submitted audio is processed by the Site's servers for analysis and deleted immediately after processing is complete. Audio data is not stored permanently or shared with third parties.

3. Acceptable Use

• You agree not to use the Site for any unlawful or malicious purpose.
• You agree not to attempt to bypass access controls, passwords, or other security measures.
• You agree not to overload, disrupt, or interfere with the Site's servers or functionality.
• Automated access (scraping, crawling, bots) is not permitted except for search engine indexing.

4. Access Control

• The Site may require passwords to access certain pages.
• The Site administrators reserve the right to restrict access to the Site or any part of it at any time, including reserving the Site for internal use.
• Temporary passwords may expire without notice.

5. Privacy

• The Site uses cookies for session management and access control.
• Anonymous usage analytics may be collected to improve the service.
• Audio submitted through the demos is processed for the stated purpose, deleted immediately after processing, and not shared with third parties.

6. No Warranty

• The Site and its demos are provided "as is" without any warranty of any kind.
• Results produced by the demos (e.g., speaker identification, fake speech detection) are estimates and should not be relied upon for legal, commercial, or security decisions.
• The Site does not guarantee availability, accuracy, or reliability of any feature.

7. Limitation of Liability

• The Site operators shall not be liable for any direct, indirect, incidental, or consequential damages arising from the use of or inability to use the Site.
• You use the Site at your own risk.

8. Intellectual Property

• The Site, its software, and its content are protected by intellectual property rights.
• You may not copy, redistribute, or republish Site content without permission.

9. Changes to Terms

• These terms may be updated at any time. Continued use of the Site after changes constitutes acceptance of the updated terms.

10. Contact

For questions regarding this agreement, please contact the Site administrators through the About page.

This demo lets you discover celebrities and public figures whose voices are most similar to yours using three separate datasets: VoxBlink2 (containing natural, in-the-wild speech recordings from 38,000 distinct individuals curated from YouTube), VoxCeleb (recordings from 5,994 distinct celebrities collected from YouTube), and the Finnish Parliament dataset (recordings from 198 members of parliament between 2009 and 2023).

To evaluate the speaker identification system, you can upload or record speech from individuals included in any of these datasets. You can also experiment with shorter speech segments to explore how much audio is needed for reliable speaker recognition.

In this demo, you can enroll and recognize the voices of up to 8 speakers.

The analysis window size can be adjusted between 500 ms and 3000 ms, which determines the length of the speech segment used to compute each data point in the result plot. Larger analysis windows generally produce more stable recognition scores but increase system latency. Smaller windows provide faster updates but may result in greater score variability.

This demo features an interactive 3D speaker map that visualizes similarity relationships between different voices. You can populate the map with speakers from the Finnish Parliament dataset, record or upload your own custom voice, and observe its placement in the 3D space. The graph dynamically draws colored spring lines connecting your voice to the closest parliamentary matches based on speaker similarity embeddings.

This demo allows you to compare up to 4 voices side by side. You can select speakers from the Finnish Parliament dataset or provide your own audio recordings. The analysis includes pairwise speaker similarity scores, F0 contour visualization, formant track visualization (F1–F3), long-term average spectrum (LTAS) visualization, and the following metrics for each voice:

• F0 median — fundamental frequency median
• F0 standard deviation — fundamental frequency variability
• F0 range — fundamental frequency range
• Jitter — cycle-to-cycle frequency irregularity
• Shimmer — cycle-to-cycle amplitude irregularity
• Harmonic-to-noise ratio — breathiness/voice quality
• Hammarberg index — spectral energy distribution
• Voicing rate — voiced segments per second
• Voiced portion — percentage of voiced speech

This demo generates speech from text. You can synthesize speech in three modes:

• Random Voice: Generate speech with a random synthetic voice.
• Voice Clone: Clone a voice from a reference audio sample.
• Voice Design: Describe the desired voice characteristics with text, and the model will attempt to match them.

This tool analyzes speech to estimate whether a recording is genuine or synthetic. It provides an overall artificiality score, along with time-resolved scores computed from 4.1-second analysis windows with a 2.05-second overlap.

This page allows you to clone a reference voice and immediately verify the quality of the generated audio. For both the original reference audio and the generated synthetic clone, you can see how the speech is rated by integrated speaker verification and anti-spoofing systems, providing automated feedback on both speaker similarity and voice authenticity.

The celebrity voice search and real-time speaker identification demos employ the RedimNet2 B2 model.

Implementation: https://github.com/PalabraAI/redimnet2

Publication: https://arxiv.org/abs/2603.11841

The text-to-speech demo employs the VoxCPM2 model.

Implementation: https://github.com/openbmb/VoxCPM

Publication: https://arxiv.org/abs/2509.24650

The spoofing detector uses the DF-Arena 500M V1 anti-spoofing model from Speech-Arena.

Model: https://huggingface.co/Speech-Arena-2025/DF_Arena_500M_V_1

Publication: https://arxiv.org/abs/2509.02859

The voice analysis demos use the SwiftF0 model for fundamental frequency (F0) estimation. It is a lightweight convolutional neural network that converts audio into a spectrogram and estimates pitch with high accuracy and speed. The model has only 95,842 parameters and runs approximately 42x faster than CREPE on CPU.

Implementation: https://github.com/lars76/swift-f0

Publication: https://arxiv.org/abs/2508.18440

The voice analysis demos use the FormantsTracker model for formant frequency (F1–F3) tracking. It is a convolutional neural network with a shared encoder and multiple decoders that predicts formants using probabilistic heat-maps rather than traditional regression, improving generalization across different speakers and speech domains.

Implementation: https://github.com/MLSpeech/FormantsTracker

Publication: https://arxiv.org/abs/2206.11632

The voice analysis demos compute several voice quality metrics using two established speech analysis tools:

Parselmouth (a Python interface to Praat) is used to compute jitter (local), shimmer (local, dB), and harmonic-to-noise ratio (HNR). These metrics are derived from a point process representation of voiced speech, with pitch estimates provided by SwiftF0.

openSMILE is used to extract additional functionals from the eGeMAPSv02 feature set, including jitter, shimmer, HNR, the Hammarberg index, and speech rate (voiced segments per second).

Parselmouth: https://github.com/YannickJadoul/Parselmouth

openSMILE: https://audeering.github.io/opensmile/

The VoxBlink2 dataset is a large-scale, audio-visual speaker recognition corpus containing millions of natural speech segments collected from YouTube videos. In its original form, it features over 110,000 distinct speakers and realistic acoustic environments.

Website: https://voxblink2.github.io/

Publication:
https://arxiv.org/abs/2407.11510

The VoxCeleb datasets are large-scale speaker identification datasets containing over a million utterances from celebrity voices extracted from YouTube videos.

Website: https://www.robots.ox.ac.uk/~vgg/data/voxceleb/

Publications:
https://www.robots.ox.ac.uk/~vgg/publications/2017/Nagrani17/nagrani17.pdf
https://www.robots.ox.ac.uk/~vgg/publications/2018/Chung18a/chung18a.pdf
https://www.robots.ox.ac.uk/~vgg/publications/2019/Nagrani19/nagrani19.pdf

The Finnish Parliament dataset is a self-collected dataset from the Finnish parliament plenary session broadcasts, featuring recordings from over 500 members of parliament.

Website: https://verkkolahetys.eduskunta.fi/fi/taysistunnot

Member Photos: https://www2.eduskunta.fi/FI/kansanedustajat/Sivut/edustajakuvat.aspx

• Javascript ES10
• Python web framework: Flask
• Front-end framework: Bootstrap 5
• Waveforms and spectrograms: WaveSurfer.js
• Icons: Ionicons
• Text polishing: ChatGPT
• Agentic coding: Devin Desktop

• Machine learning framework: Pytorch
• Interprocess communication: ZeroMQ
• Scientific computing: NumPy
• Audio functions: Librosa, Torchaudio

• Flask tutorials: YouTube link
• Animated text tutorial: YouTube link

Ville Vestman, PhD

Website: https://cs.uef.fi/~vvestman/

Google Scholar: https://scholar.google.com/citations?user=aPZBcWgAAAAJ

Professor Tomi H. Kinnunen, PhD

Website: http://cs.joensuu.fi/pages/tkinnu/webpage/

Google Scholar: https://scholar.google.fi/citations?user=e3SPjpoAAAAJ

Generalized Voice Anti-Spoofing and Voice Biometrics (SPEECHFAKES), Academy of Finland (09/2022 – 08/2026).

Computational Speech Group at the School of Computing, University of Eastern Finland.

The maximum duration for recording speech is set to either 30, 60, or 120 seconds, depending on the specific demo. Imported audio files are clipped to a maximum duration of 3 minutes. If the duration of the imported audio exceeds the maximum limits mentioned earlier, only a shorter portion can be selected for analysis, rather than the entire file.

The VoxCeleb dataset contains a number of labeling errors. In cases where there is a discrepancy between the name and the voice, the application compares the input speech to the voice in the video, rather than the speaker indicated by the name. Labeling errors are less prevalent in the parliament dataset.

Ensure your mic is not muted. Check whether the log at the bottom of the page contains helpful information for troubleshooting and resolving the issue.

Sometimes refreshing the page fixes the problem.

The frequency spectrum visualizer becomes active when the microphone is functioning properly.

• Added user agreement.
• Added admin page.
• Added Analyze & Compare demo.
• Added microphone permission help dialog with instructions for re-enabling access after denial.
• Improved server stats reporting with offline mode detection.

• Upgraded the speaker identification model to ReDimNet2-B2.
• Added the new VoxBlink2 speaker recognition dataset.
• Introduced the new Speaker Graph and Voice Clone & Verification demos.
• Added the new text-to-speech demo using the VoxCPM2 model.
• Removed the old speech synthesis demo based on SpeechT5 and the old voice conversion demo based on FreeVC.
• Updated the anti-spoofing demo model to DF-Arena 500M V1.
• Redesigned the entire UI with a unified neon visual theme.
• Expanded the audio visualizer with toggleable view modes. The original frequency spectrum is now joined by new waveform and spectrogram views.
• Added speaker images for the parliament dataset, with audio playback available by clicking the images.
• Removed the gauge.js, js-cookie, jQuery, and Rocher font dependencies.

• Bug fix: when a new audio file was loaded, the computation was done based on the selection of previous audio file, if some speech segment was selected.
• Updated anti-spoofing visuals.
• Updated anti-spoofing model.

• Automatically change http to https to prevent recording problems.
• Version history added.

Website ready.

Development began.