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.

Original post →

Open-Source Voice Model Indexes 174 Hours of Apollo Audio

More from Multimodal