XNews

Audio AI: Suno data breach and ASR tool developments

Audio AI: Suno data breach and ASR tool developments

Suno's training data and user information were breached. - An attacker used a worm named Shai-Hulud to obtain employee credentials and access Suno systems. - Stolen materials included training-related source code and a dataset scraped from YouTube Music, Pond5, Deezer. - Reported scraped volumes: 113,879 hours from YouTube Music, 62,117 hours from Pond5, 12,287 hours from Deezer. - Suno reportedly planned to download about 1,000,000 hours of podcasts. - The breach exposed Suno user emails, phone numbers, and Stripe-related payment information for tens of thousands of users; some card details were reportedly included. - Separately, Qwen3 ASR plus a Qwen3-ForcedAligner achieved accurate word-level timestamps using a 0.6b model with modest local resources. - Pyannote and WeSpeaker were noted as open-source options for speaker recognition; cloud models (example: Doubao on Volcano Engine) were cited for higher-quality transcription.

Key insights

Suno breached; training code and large scraped audio corpus accessed

A security incident at Suno involved an attacker using a worm named Shai-Hulud to obtain employee credentials and access company systems. The intruder retrieved training-related source code and lists of scraped audio content, including 113,879 hours from YouTube Music, 62,117 hours from Pond5, and 12,287 hours from Deezer. Reporting indicated Suno had plans to download roughly 1,000,000 hours of podcasts. The breach also exposed user contact details and payment-related records.

Breached data includes user emails, phones, and payment info

The intrusion at Suno reportedly exposed user email addresses and phone numbers, along with payment details linked to Stripe. Reporting noted that several hundred thousand users were affected and that some card information appeared in the materials accessed during the breach.

Qwen3 ASR plus ForcedAligner yields precise timestamps with small model

Testing of Qwen3 automatic speech recognition with a Qwen3-ForcedAligner model produced accurate word-level timestamps. The evaluator reported that a 0.6b Qwen3 model delivered good transcription and forced-alignment accuracy while running locally with modest resource usage. This configuration addressed typical issues such as misaligned timestamps in previous tools.

Speaker recognition and cloud services cited as complementary solutions

Open-source speaker recognition tools Pyannote and WeSpeaker were cited as options for speaker identification, with limitations when many speakers overlap. The report noted that higher-quality transcription and faster processing remain available via paid cloud offerings, with an example given of a Doubao recognition model on Volcano Engine as a commercial alternative.

Source posts