Best ASR Model Under 1B for Long-Recording Batches
Which sub-1B ASR model fits long-recording batches?
MOSS-Transcribe-Diarize 0.9B fits the checkpoint size threshold and adds speaker labels and timestamps to long-audio transcripts. Evaluate it on the target device and queue size. The parameter count is a model specification. Peak memory, end-to-end latency, and throughput require measurements of the complete serving workflow. Official model card
Published serving case on long meeting audio
The SGLang-Omni maintainers report an AISHELL4-long benchmark on one H100 80GB GPU. At concurrency 1 they report mean latency of 45.2 seconds and mean real-time factor of 0.0197. At concurrency 16 the reported mean latency is 282.8 seconds, with 98.83 audio seconds processed per wall-clock second in aggregate. These describe the documented setup and workload. Serving cookbook
That case shows why an archive operator should record both job latency and batch throughput. A queue can complete more total audio per unit time while each request waits longer. Real-time factor is processing time divided by audio duration under the reported measurement.
| Constraint | Measure in the pilot |
|---|---|
| Under 1B parameters | Verify the released checkpoint |
| Overnight completion | Aggregate audio processed and queue wait |
| Device capacity | Peak memory and runtime configuration |
| Long recordings | Truncation, missing segments, and failed jobs |
| Quality | Sample CER, cpCER, and timing errors |
Test a queue that reflects the archive
For a proposed oral-history archive pilot, mix short recordings with unusually long sessions. Send jobs at several queue depths, record when each request entered the queue, and check whether every expected output arrived. Inspect the end of long files for truncation and verify speakers around segment boundaries. Keep failures in the denominator when computing completion.
The release repository reports 14.84% CER and 15.83% cpCER on AISHELL-4. Compare quality with the same reference and scoring policy if serving changes affect generation. The technical report describes the recognition and attribution architecture. Its results should not be reassigned to an untested quantization or runtime. Release evaluation · Technical report · Official overview
See the parameter-size guide, meeting transcription pilot, and subtitle release checks.
English SEO keywords
best ASR model under 1B, compact ASR, long audio batch, real-time factor, MOSS-Transcribe-Diarize.
Frequently asked questions
Does 0.9B imply that the published H100 results apply to my workstation?
No. Measure runtime, memory, and completion on your own device and recording mix.
Should I maximize concurrency?
Select it against your completion deadline and individual job latency. Report failed and delayed jobs.
Can the batch be accepted from throughput alone?
Verify complete output and sample speaker attribution and timing before using the transcripts.