Skip to content
Last updated

AI Voice Watermarking

Dixa's AI voice agent produces synthetic speech on calls. Every vocalization it generates carries an inaudible watermark, so that the audio can later be identified as AI-generated rather than human.

This page is for anyone who needs to establish whether a Dixa call recording contains AI-generated speech. You can upload a recording to get a signed report, and verify a report you have received.

Watermarking supports the transparency obligations that apply to AI-generated audio under Article 50 of the EU AI Act. It covers Dixa's AI voice agent only, and not Dixa's other AI features.

This page explains how Dixa's watermark works and how to check for it. It's not legal advice on your own obligations under the EU AI Act or any other law. Confirm those with your legal or compliance team.

Check a recording

Upload a call recording to get a signed PDF report that you can keep as an audit record. The report records the verdict, the detection scores for each channel, and the SHA-256 hash of the file you uploaded, which identifies the recording.

Detection needs about 2 seconds of uninterrupted AI speech. In a stereo recording, the watermark is detected if either channel carries it.

What a report does and does not show

A report states whether Dixa's watermark was detected in the file you uploaded. A positive result indicates that the audio contains speech generated by Dixa's AI voice agent.

A negative result does not establish that the audio is human: the AI speech may be too short to detect, the recording may have been degraded by compression, transcoding or editing, or the audio may come from a source that Dixa does not watermark.

Reports are provided for information and are not a certification or a compliance attestation.

Before you upload

Call recordings usually contain personal data. By uploading a recording, you confirm that you are entitled to share it with Dixa for this purpose, that you have a lawful basis for the processing, and that your recording disclosures cover this use.

Dixa processes the file in the EU solely to generate the report and deletes it immediately afterwards. The report is returned to your browser and is not retained by Dixa. For Dixa customers, this processing falls under your existing data processing agreement.

16-bit PCM WAV, up to 110 MB.

Supported format

PropertyRequirement
File typeWAV
Encoding16-bit PCM
Sample rate8000, 16000, 44100 or 48000 Hz
Channels1 (mono) or 2 (stereo)
LengthAt most 10 minutes
File sizeAt most 110 MB

Many telephony systems save call recordings as WAV files with 8-bit µ-law or A-law encoding, which is not supported. Convert those to 16-bit PCM first.

Convert a recording

You can convert most recordings with FFmpeg. To see what format a recording is in, run:

ffprobe -v error -select_streams a:0 -show_entries stream=codec_name,sample_rate,channels:format=duration -of default=noprint_wrappers=1 recording.wav

A supported recording shows codec_name=pcm_s16le, one of the supported sample rates, 1 or 2 channels, and a duration of at most 600 seconds.

To convert a recording in another format, such as MP3 or a µ-law WAV, to 16-bit PCM WAV while keeping its sample rate and channels, run:

ffmpeg -i recording.mp3 -c:a pcm_s16le recording.wav

If the sample rate is not supported, add -ar with a supported rate:

ffmpeg -i recording.mp3 -c:a pcm_s16le -ar 16000 recording.wav

If the recording is longer than 10 minutes, split it into 10-minute parts and upload each part separately:

ffmpeg -i recording.wav -f segment -segment_time 600 -c:a pcm_s16le part-%03d.wav

AI speech that falls right at the boundary between two parts may be too short to detect in either part.

Verify a report

Every report is digitally signed by Dixa. Upload a report to check that Dixa signed it and that it has not been changed since. Any change after signing, including added annotations or form fields, makes verification fail.

PDF, up to 16 MB.

Dixa signs reports with its own certificate rather than one issued by a certificate authority your PDF viewer trusts. If you open a report in Adobe Acrobat Reader or another viewer, it will show the signature as intact but the signer's identity as unknown. This is expected. The form above is the reliable way to confirm that a report comes from Dixa and has not been altered.

This page explains how Dixa's watermark works and how to check for it. It is not legal advice on your own obligations under the EU AI Act or any other law. Confirm those with your legal or compliance team.