How to spot fake audio in the age of AI

On June 5, 2025, an audio clip began circulating in Uganda, allegedly featuring well-known journalist Solomon Serwanjja in a private conversation. Many shared the clip, while others doubted its authenticity. Was it real or generated by artificial intelligence?

Some people immediately questioned the authenticity of the clip, calling it a deepfake. One person argued that the voice sounded too robotic and too clean, lacking the natural pauses and breaths you would expect in a real conversation. Others stated that without access to the original audio file, it is almost impossible to ascertain its authenticity. The discussion was largely driven by speculation and sensationalism on social media, where unverified claims often go unchecked.

If the audio is real, it could represent a serious invasion of Serwanjja’s privacy. If it is fake, it could be one of the first known instances of AI-generated disinformation targeting a journalist in Uganda. Either way, the incident points to a troubling new reality. We have seen fake news headlines, doctored images and impersonator accounts before. Now, we also have to contend with convincing audio that sounds just like someone you know or trust.

What is a deepfake, anyway?

A deepfake is a type of synthetic media created using artificial intelligence to mimic the appearance, voice or behaviour of a real person. It is typically made using machine learning algorithms that are trained on large datasets of a person’s voice or face, allowing them to generate realistic imitations. In the case of a voice deepfake, AI technology is used to replicate someone’s unique speech patterns, accent, tone and even emotional inflections, making it sound as though they said something they never actually did.

Voice deepfakes are particularly concerning because they are easier and faster to create than video deepfakes, and are often more difficult for listeners to scrutinise, especially in casual listening environments. They can be convincing even with short audio samples, and their impact is amplified on platforms like WhatsApp or X (formerly Twitter), where audio quality may already be degraded or compressed. This makes it harder for ordinary users to determine whether the voice is genuine or artificially generated.

As AI tools become more accessible and sophisticated, the line between real and fake may become increasingly blurred. This poses serious challenges not just for individual privacy, but also for journalism, law enforcement and public trust.

How can you tell if an audio leak is real?

You do not need to become an AI expert, but you should know how to ask the right questions and use available tools when confronted with a suspicious leak. Here are some steps anyone can take to verify if a piece of audio is real or fake:

• Check the origin – Who posted the audio? Is there a clear chain of custody? Was it accompanied by screenshots or text claiming its context?
• Examine the audio – Use free tools (see list below) to view the audio waveform. Do you see unnatural gaps or repeated patterns? In higher-end programmes, you can check spectrograms (visual sound maps) for missing frequencies or signs of artificial voice generation. Listen for unusual cadence, lack of emotion or speech that sounds oddly clipped.
• Seek human clues – Play the audio to people familiar with the speaker’s real voice; do they recognise it? Are there inconsistencies in the language, accent or phrasing? Does the emotional tone match the context? Does the speaker sound too calm or too detached?
• Reach out – If possible, contact the person allegedly in the audio. In this case, has Solomon Serwanjja confirmed or denied it? His public statement, or silence, is part of the verification picture.
• Look for supporting or contradictory evidence – Are there messages, videos or photos that back up the events described? Are there eyewitnesses or other records that support or challenge the story?

Tools to help you spot deepfake audio

Here are some free audio tools and methods you can use to help with this task. But remember: no single tool can give you a definitive answer. Each method can point to red flags, but it is only by combining technical analysis, contextual evidence and human judgment that can you build a reliable conclusion. Treat these tools as part of a broader verification process, not a final verdict.

Online AI audio detectors

While not always 100% foolproof, several online tools are emerging that claim to detect AI-generated voices. These often use machine learning to identify patterns in speech that are characteristic of synthetic voices.

• DeepFake-o-meter (University at Buffalo): This is an open platform that integrates various state-of-the-art methods for AI-generated image, video and audio detection. It is more research-oriented but could be a powerful tool for those willing to dive into it.
• AI Voice Detector: This tool claims to detect AI-cloned voices from major AI models and even has a Chrome extension. Note that some online detectors like this one might have limitations on file size or require a subscription for full features.
• Resemble Detect: While Resemble AI is a voice generation platform, it also offers “Resemble Detect” which is an AI deepfake detection tool. It claims high accuracy in identifying manipulated audio.

Audio analysis software

• Sonic Visualiser (free, open-source): Designed for detailed visualisation, analysis and annotation of audio recordings. It offers highly configurable views, and can be useful for musicologists, archivists and signal-processing researchers. This allows for a deeper dive into the frequency spectrum and could reveal subtle anomalies.
• WASP (free): WASP from University College London is designed for introducing students to acoustic analysis. It is light on disk space and user-friendly for basic acoustic examination.
• SIL Speech Analyzer (free): Offers more functions than WASP with a flexible display format and good pitch-tracking. It can be useful for detailed phonetic analysis of suspected AI audio.

General audio editors (for quick checks and manipulations)

• Audacity (free, open-source): This is a widely popular and powerful multi-track audio editor and recorder. It is excellent for quick visual inspection of waveforms, basic cuts, noise reduction, and applying various effects. Its ability to display the sound wave helps in spotting unnatural edits or repeating patterns, making it a foundational tool for initial anomaly detection.
• WaveSurfer (free, open-source): Another sound visualisation and manipulation tool that is user-friendly for both novice and advanced users. It can be adapted for various tasks, including speech or sound analysis and annotation, offering good visual feedback on audio structure.
• Ocenaudio (free): A fast and easy-to-use audio editor that is great for daily audio editing tasks. It operates across platforms and features real-time preview of effects, making it quick to apply and test changes. Its clean interface makes it accessible for quick checks.
• LMMS (Linux MultiMedia Studio) (free, open-source): While primarily a digital audio workstation (DAW) for music production, LMMS includes a sample editor that allows for basic waveform manipulation, trimming and applying effects to individual audio clips. It is more geared towards creation but can be used for simple analysis tasks if you are already familiar with DAWs.
• VLC Media Player (free, open-source): While not a dedicated editor, VLC can play almost any audio format and has basic playback controls that can be useful for quick listening tests. Its ability to adjust playback speed can sometimes reveal subtle unnaturalness in AI-generated speech that might be missed at normal speed.

Even without technical tools, training your ear to listen for unnatural intonation, unusual pacing or odd pronunciations can be effective. AI voices are getting better, but sometimes they still sound “too perfect” or lack the subtle imperfections of human speech.

Ethical considerations for the media

Whether or not the audio in question is genuine, journalists and editors have a responsibility to consider the potential consequences of publishing it. Even if the story is true, releasing it without careful thought can lead to unnecessary harm. Ethical journalism demands reflection on whether the public truly needs to hear the content.

Publishing personal audio, particularly without consent, also raises serious legal concerns. In Uganda, this can breach the Data Protection and Privacy Act (2019) and could also result in legal challenges under defamation or cybercrime laws.

It is therefore important for media organisations to adopt a careful and principled approach in publishing content like this.

• Always verify the material before publishing and do not run with a leak unless there is reasonable confidence in its authenticity.
• Assess whether there is a clear public interest. If the audio contains sensitive or irrelevant personal details, redact them.
• Before publication, offer the subject a right of reply, allowing them to respond to the claims or context.
• Finally, if something has not been fully verified, make that clear to the audience.