Open-Source Voice Model Indexes 174 Hours of Apollo Audio

vanstriendaniel · x · 2026-07-11

The post introduces **MOSS-Transcribe-Diarize**: a **0.9B** parameter, **Apache 2.0** licensed open-source model capable of performing **transcription, speaker diarization, and timestamping** in a single inference pass. ### Real-World Use Case The author used it to make **174 hours** of Apollo 11 mission audio searchable, resulting in: - Searchable results organized by "who said what and when" - A total of **45,355** timestamped speaker segments - A processing cost of **$9.46** ### Performance & Deployment The audio came from the internet archive, and the entire batch was processed via a Hugging Face Job: - Runtime environment: **a100-large** - Total time: **3.8 hours** - Throughput: **47x realtime** - The model was served by **sglang** inside the Job The original post also detailed implementation notes: they run short-sequence, long-sequence, and multi-speaker tests on every `sglang-omni` commit to ensure performance and correctness via continuous regression testing. They also explained their choice of server-based serving over an in-process engine, as the server side more easily captures throughput gains from continuous batching.

Related event: Open_MOSS Releases Apollo Audio Transcription Model(2 posts)→

Original post →

More from Multimodal

Multimodal channel →