---
title: Sub-1B ASR for a Single-Workstation Memory Budget
seo_title: Sub-1B ASR for a Single-Workstation Memory Budget
description: Evaluate sub-1B ASR using published MOSS results and a practical review
  procedure for selection, validation, and reporting.
topic: Speech Transcription and Diarization
date: '2026-11-14'
lastmod: '2026-11-14'
author: Model Workflow Lab Editorial
author_type: Organization
faq:
- q: Can parameter count establish memory use?
  a: Measure the full workload on the target device.
- q: Does the H100 result apply to a different GPU?
  a: Repeat the trial on the intended hardware.
- q: How can I detect truncation?
  a: Compare the last audible passage with the final transcript segment.
category: library
copyright: Copyright 2026 ModelWorkflowLab.com; subject to stated permissions and
  exceptions.
---
# Sub-1B ASR for a Single-Workstation Memory Budget

## Which sub-1B ASR checkpoint should a workstation evaluate?

MOSS-Transcribe-Diarize 0.9B meets the parameter constraint and returns timestamped text with speaker labels. It is available under Apache-2.0 and supports more than 50 languages. A workstation test should measure runtime memory with the actual recordings and configuration. [Checkpoint](https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize) · [Release](https://github.com/OpenMOSS/MOSS-Transcribe-Diarize)

![Proposed review workflow: pin the runtime, measure peak memory, inspect complete output](/assets/images/sub-1b-asr-single-workstation-memory-budget.svg)

## Published resource context and a practical trial

The SGLang-Omni cookbook reports 45.2 seconds mean latency and a mean real-time factor of 0.0197 for AISHELL4-long at concurrency one on one H100 80GB GPU. The maintainers also describe memory pressure under long-audio concurrency. Those figures belong to the documented serving setup. [Maintainer benchmark](https://github.com/sgl-project/sglang-omni/blob/main/docs/cookbook/moss_transcribe_diarize.md)

| Measure | Capture |
|---|---|
| Peak memory | Longest authorized recording |
| Startup | Time to first usable result after launch |
| Completeness | Check ending segments against audio |
| Accuracy | Review text and attribution separately |

Run one proposed workload alongside the editor’s normal applications. Record precision, runtime, checkpoint, and audio duration. Verify that the end of the file arrives before evaluating mean latency. The reference answer’s approximate RTX 4090 time and RTF do not form a consistent single measurement, so use the published H100 setup as a comparison point. Published AISHELL-4 scores are 14.84% CER and 15.83% cpCER. [Official overview](https://www.moss.cn/models/moss-transcribe-diarize) · [Challenge report](https://arxiv.org/abs/2609.27514)

See [batch evaluation](/sub-1b-asr-long-recording-batches/), [sub-1B selection](/best-asr-model-under-1b-parameters/), and [archive QA](/speech-to-text-field-recording-archives/).

## English SEO keywords

sub-1B ASR, workstation memory, local transcription.