Show HN: Trelis Tiron – open-weights multi-speaker meeting transcription (huggingface.co)

Trelis Research has announced Tiron, an advanced multi-speaker meeting transcription model capable of transcribing and attributing speech to up to eight speakers in real-time. Released on July 21, 2026, Tiron operates on the Whisper large-v3 architecture, leveraging an extended token vocabulary and producing inline transcripts complete with speaker turn markers and timestamps every 30 seconds. Its deployment as a drop-in replacement for WhisperForConditionalGeneration underscores its utility, particularly for whole-meeting transcriptions, where it outperformed its closest competitor, AssemblyAI's universal-3-pro, across several test sets, showcasing reductions in concatenated-permutation word error rate (cpWER). This release is significant for the AI/ML community as it addresses critical challenges in accurately transcribing long meetings with multiple speakers, showcasing a notable advancement in the field of automatic speech recognition (ASR). Tiron's design facilitates stable speaker identification through ECAPA voice embeddings and cross-window linking, which are crucial for maintaining continuity across longer discussions. Developers and researchers can integrate Tiron easily using the open-source harness available on GitHub, promoting further innovations in dialogue systems and contributing to enhanced usability in professional environments.