GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Arabic-STT
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Arabic-STT
Model Tree
Qwen/Qwen3-ASR-1.7B
Input Modalities
Output Modalities
Supported Functionality
text
model = run6b + 0.6 x (run10 - run6b) + 0.6 x (avg - run6b)| checkpoint | |
|---|---|
base run6b | Arabic-STT/qwen3-asr-1.7b-ar-run6b |
run10 | Arabic-STT/qwen3-asr-1.7b-ar-run9, tag run10-rl-s750 (commit e938b6d), GSPO from run9b |
avg | the 0.5 / 0.5 weight average of sip-v5 and sip-v5-grpo1 |
Merged tensor by tensor in float32, saved in bf16; config, tokenizer and processor from run10. merge.json holds the
recipe.
bf16, greedy decoding, through the qwen3-asr-arabic-api server (long files cut at pauses into chunks of at most 30 s). Word error rate in %, the same normalizer applied to reference and hypothesis.
| set | base Qwen3-ASR-1.7B | this model |
|---|---|---|
| Audioluna call samples (20 calls, Arabic + English; mean of 3 runs, normalize_ext) | 22.01 | 12.35 |
| Audioluna, plain scoring (lower-case, punctuation removed) | - | 13.40 |
| Test Arabic (505 segments) | 18.55 | 9.87 |
| SIP call sample (held out) | 45.58 | 13.86 (9.16 with every digit counted as a word) |
| FLEURS ar_eg | 8.98 | 4.64 |
| Casablanca Morocco / Egypt / Jordan / UAE | 85.80 / 55.66 / 35.80 / 48.76 | 37.56 / 27.16 / 19.04 / 32.06 |
| SADA (Saudi) | 44.27 | 25.12 |
| Common Voice 17 Arabic | 17.72 | 5.91 |
| Mixat (Emirati Arabic-English code-switching) | 25.04 | 15.05 |
| AMI / Earnings22 | 18.53 / 10.22 | 14.92 / 9.21 |
| LibriSpeech clean / other | 1.60 / 3.73 | 1.72 / 4.77 |
| GigaSpeech / SPGISpeech / VoxPopuli / Common Voice English | 8.10 / 2.72 / 5.52 / 7.44 | 9.12 / 3.77 / 6.01 / 9.81 |
The public sets are 250 random test utterances each (Mixat 198; seed 0), scored with Whisper's English normalizer for
English and normalize_ext for Arabic. The model is much better than the base model on Arabic, dialects, code-switching
and calls (13 of 19 sets, paired bootstrap 95 %) and 0.5-2.4 points worse on clean read English (5 sets).
Caveat: the merge and its weights (0.6 / 0.6) were chosen on Audioluna, so the Audioluna number is slightly optimistic; the other sets were not used for any choice.
No system prompt and the model's own language detection give the default written format (numbers as digits, English
in Latin script). Loadable with from_pretrained (bf16 weights + processor) and with vLLM. Long recordings: cut at
pauses into chunks of at most ~25-30 s; two-channel calls summed to mono.