Google Gemini 3.5 Transcribe Models Differ Sharply on Speaker Diarization Support

·1 views

A developer building a macOS meeting translation app discovered that Google released two similarly named models — gemini-3.5-transcribe-live and gemini-3.5-transcribe — with significantly different capabilities. The live streaming version uses WebSocket via the Live API and supports real-time interim transcription, but does not support speaker diarization or word-level timestamps. Speaker diarization, which identifies who said what, is only available in the non-streaming version, which requires uploading a complete audio file after recording and supports up to eight speakers. This architectural limitation meant the developer could not achieve real-time speaker identification and had to design the app around a post-meeting upload workflow for diarized transcripts. The live model does offer features like automatic language detection, a SMART cleanup mode that removes filler words, and a custom vocabulary option supporting up to 1,000 technical terms.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Untested Backups Are Unreliable: Why Restore Drills Are Essential

A cautionary piece by IT professional Serguey Shinder highlights the critical flaw in assuming backups work simply because jobs report success. Shinder recounts a real incident where months of seemingly successful backups produced incomplete data that was only discovered during an actual recovery emergency. The core argument is that a backup is only valid once it has been fully restored to a real system and verified for completeness. He warns that backup jobs can silently fail while still appearing healthy, creating a false sense of security. Shinder urges IT teams to schedule regular restore tests, track recovery times, and treat the process as routine to avoid discovering failures during a live crisis.

0 comments Read more at DEV Community

ProgrammingHacker News ·

Security researchers warn that bug rumors alone now trigger active exploits

A security analysis highlights a growing trend where even unconfirmed rumors of a software vulnerability are sufficient to prompt attackers to search for and develop exploits. This shift means that the window between public speculation and active exploitation has effectively collapsed. Attackers are now treating vague threat intelligence as a starting signal rather than waiting for official disclosures or proof-of-concept code. The development raises serious concerns about how organizations respond to unverified vulnerability reports. It underscores the need for faster patching cycles and more cautious handling of early-stage security information.

0 comments Read more at Hacker News

ProgrammingHacker News ·

Nvidia Confident It Can Sustain Revenue Growth Amid AI Infrastructure Surge

Nvidia has expressed confidence in its ability to continue generating strong revenue to support the ongoing artificial intelligence boom. The chipmaker, which has seen explosive growth driven by surging demand for AI hardware, insists its financial momentum is sustainable. Nvidia's graphics processing units have become central to AI data center buildouts by major tech companies worldwide. The company's statements come as investors and analysts closely watch whether AI infrastructure spending will remain at elevated levels.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Most Cyberbreaches Stem from Poor Hygiene, Not Sophisticated Attacks

Security expert Serguey Shinder argues that the majority of real-world data breaches are caused by basic oversights rather than advanced hacking techniques. Common culprits include reused passwords, unpatched servers, misconfigured storage buckets, and employees falling for phishing emails. Shinder notes that organizations often prioritize flashy security tools over unglamorous but essential practices like multi-factor authentication, asset tracking, and timely access revocation. He warns that attackers rarely need to be sophisticated when fundamental security hygiene is neglected. According to Shinder, closing obvious gaps consistently is what genuinely prevents breaches, not chasing complex threats.

0 comments Read more at DEV Community