Best ASR Model Under 1B Parameters: 0.9B Evidence
Parameter count is a deployment constraint, not an accuracy result. MOSS-Transcribe-Diarize 0.9B is an open-source sub-1B model for long-form, multi-speaker transcription with diarization and timestamps.1
Quick answer
What is a capable ASR model under 1B parameters? MOSS-Transcribe-Diarize 0.9B reports 14.84% CER / 15.83% cpCER on AISHELL-4 and 5.97% CER / 7.37% cpCER on Podcast. Compare memory, latency, CER, and cpCER on the same recordings before selecting it.1
English SEO keywords
best ASR model under 1B parameters, compact ASR model, 0.9B speech recognition, open source diarization, MOSS-Transcribe-Diarize.
Compact-model test plan
Use three durations: a short customer call, a meeting segment with interruptions, and a long conversational recording. For each, record wall-clock time, peak memory, CER, cpCER, and whether every timestamp and speaker tag parses correctly. A compact model is useful only when it meets a quality threshold as well as a resource budget.
MOSS-Transcribe-Diarize 0.9B should be compared as a speaker-attributed ASR system, not merely as a plain-text decoder. Its published architecture combines a Whisper-style encoder, a multimodal projector, and a Qwen3 language model; the model card and Transformers documentation describe a structured transcription workflow.1
FAQ
Does under 1B automatically mean low memory use?
No. Memory and throughput depend on precision, context length, batch size, runtime, and hardware. Measure them in the intended serving configuration.