GLM-ASR — The GLM Speech Recognition Model

GLM-ASR is the speech member of the GLM model family: a dedicated automatic speech recognition model that turns audio into text in real time, with a published character error rate of 0.0717. It is not a language model with an audio front-end bolted on — it is a purpose-built recognition system, and it is evaluated on the metric that actually matters for transcription rather than on reasoning benchmarks.

A character error rate of 0.0717 means roughly seven characters in a hundred are wrong, counting insertions, deletions and substitutions. In practice that is comfortably readable output for meeting notes, captions or voice commands, with proper nouns and domain jargon remaining the usual sources of the residual error. Like the rest of the family, GLM-ASR is MIT licensed with weights published openly, so the audio never has to leave your own infrastructure.

0.0717
character error rate
published for GLM-ASR real-time recognition
MIT
open-source license
MIT, with no regional restrictions
Real-time
recognition mode
streaming speech-to-text, not batch-only

Free AI Playground — Try a Live Open-Source LLM

No sign-up, no account, no credits. Type a prompt and get a streamed answer.

Response will stream here...

This playground runs a free open-source LLM via OpenRouter. It is not GLM-5.2 and does not use GLM weights — it's here so you can test a live model while you read the published GLM benchmark data below.

How to Read a 0.0717 Character Error Rate

Character error rate is the edit distance between the transcript and the reference, divided by the length of the reference — so 0.0717 means about seven errors per hundred characters. It is a stricter and more language-agnostic measure than word error rate, which is why it is the standard metric for systems that have to handle languages without clear word boundaries. What the number does not tell you is where the errors fall. Recognition systems are consistently weakest on names, product terms, acronyms and heavy accents, and strongest on ordinary conversational speech, so a transcript can sit at 0.0717 overall while being effectively perfect on the body text and wrong on exactly the terms you wanted to search for. If accuracy on domain vocabulary matters, evaluate on your own audio and consider biasing the decoder toward a term list rather than judging by the headline figure alone.

Character error rate
0.0717
License
MIT open source
Positioning
Real-time speech-to-text
Modality
Audio → text
Weights host
HuggingFace · ModelScope

GLM-ASR and the Published Benchmark Table

GLM-ASR is a speech recognition model and does not appear in the published text-LLM comparison table. Its published figure is the 0.0717 character error rate quoted above; no LLM benchmark column exists for it, and none is invented here.

The 17-benchmark comparison table on this site covers text models — reasoning, coding and agentic evaluations — and a speech recogniser is not comparable on any of those axes. That is why this page reports the character error rate and stops there. If you need the text side of the family, the flagship page carries the full published column and the benchmarks hub carries all 17 evaluations across 8 models.

See all 17 GLM benchmarks in detail →

The Text Side of the Family, For Context

Comparison on Terminal-Bench 2.1 (Terminus-2), as published by the model authors. Shown for orientation only — these are text models, not speech models.

Speech recognition is usually one stage of a pipeline: transcribe, then reason over the transcript. If that is what you are building, GLM-ASR handles the first stage and a text GLM model handles the second, and the chart above shows the published gap between the two most recent text releases on Terminal-Bench 2.1 — 81.0 for GLM-5.2 against 63.5 for GLM-5.1. Both stages can run on your own hardware under the same MIT license.

Read the full GLM vs Claude comparison →

Where GLM-ASR Sits in the GLM Model Lineup

GLM-ASR sits outside the text lineage entirely — it is the family's speech branch rather than a step between GLM-4.7 and GLM-5. Pair it with a text GLM model when you need transcription followed by reasoning; the lightweight GLM-4.7-Flash tier is a natural partner when the whole pipeline has to run locally, and GLM-5.2 when the reasoning stage is genuinely hard.

More GLM Models

View all

GLM-ASR FAQ

What is the GLM-ASR model?
GLM-ASR is the speech recognition model of the GLM model family. It performs real-time speech-to-text transcription with a published character error rate of 0.0717 and, like the rest of the family, ships under an MIT license with open weights.
What does a 0.0717 character error rate mean?
It is the edit distance between the produced transcript and the reference divided by the reference length — roughly seven wrong characters per hundred, counting insertions, deletions and substitutions. Errors cluster on names, acronyms and domain jargon rather than on ordinary conversational speech.
Can GLM-ASR transcribe in real time?
Yes — real-time speech-to-text is its stated design point, which means it emits text as audio streams in rather than requiring a complete file first. That is what makes it usable for live captioning and voice control rather than only for offline transcription jobs.
Is GLM-ASR open source?
Yes. It is an MIT-licensed GLM model with weights published on HuggingFace and ModelScope and no regional restrictions, so you can run transcription entirely on your own hardware and keep the audio inside your network.
Does GLM-ASR appear in the GLM benchmark table?
No. The published 17-benchmark comparison covers text models on reasoning, coding and agentic tasks, which a speech recogniser cannot be scored on. Its published metric is the character error rate; for text performance see the GLM-5.2 page or the full benchmarks hub.

Try a Live AI Model Free — No Account Needed

Read the GLM model benchmarks, then put a real model to work. The playground above is free and needs nothing from you; if you want a fuller AI toolkit, our partner's free tier starts here.