Quivra 1.1 (fine-tuned Cohere Transcribe Arabic 07-2026, sherpa-onnx int8)

Quivra 1.1 (2026-11-13), quivra-arabic-1.1-int8.tar.zst: Quivra 1.0 trained further with LoRA (encoder and decoder, rank 16, 4 epochs) on the Mixat training split (Emirati Arabic–English, ~5 h after removing clips whose transcript does not cover the audio; spelling repaired: final ة/ي, numbers as digits), the ArzEn training split (Egyptian Arabic–English, 5.5 h), the 1.0 speaker's recordings and 170 silence/noise clips with empty transcripts. Held-out results against Quivra 1.0, normalized WER, model output only: owner's recordings 26.5% → 18.3%, Mixat test (200 clips) 19.8% → 14.8%, ArzEn validation 6.2% → 6.4%; English words kept in Latin letters 62% → 94% and 40% → 84%; text invented on silence/noise 5 → 0 of 18 clips; same speed. The archive holds the four files below; the loose files in this repository are still Quivra 1.0 (for older app versions).

Quivra 1.0 (2026-11-12): fine-tuned with LoRA (decoder, rank 16) on ~32 minutes of Levantine Arabic / English code-switched speech read by one speaker, plus English and MSA replay, then merged and converted. English words come out in Latin letters more often. On held-out real recordings (not used for training) the normalized WER went from 17.7% (base + Quivra's loanword step) to 16.2%; it is tuned on one voice, so other voices may gain less. Version 1 (the plain conversion of the base model) is in this repository's history.

An int8 ONNX conversion of CohereLabs/cohere-transcribe-arabic-07-2026 for sherpa-onnx (Cohere Transcribe model type), made for the Quivra dictation app by Fox Labs. Same file layout and interface as sherpa-onnx's own conversion of the 14-language base model.

• encoder.int8.onnx + encoder.int8.onnx.data: Conformer encoder; outputs the cross-attention K/V for the decoder.
• decoder.int8.onnx: decoder with self-attention KV cache.
• tokens.txt: vocabulary (identical to the base model's).

Quantisation: ONNX Runtime dynamic int8 (MatMul weights). Tell the model the language (ar); English words spoken inside Arabic are often written in Arabic letters, so a post-processing step can help.

Licence and credit

Cohere Transcribe Arabic © Cohere, licensed under the Apache License 2.0. Converted to ONNX and int8 by Fox Labs; the fine-tune is Fox Labs' work under the same licence. Use it under the original Apache-2.0 terms.

Quivra 1.1 training data: Mixat (Al Ali et al., MBZUAI; CC BY-NC-SA 4.0) and ArzEn (Hamed et al.; released for research). Their terms apply to the use of this fine-tune.