cohere-cs-decoder-middle

nsa01n/cohere-cs-decoder-middle

Decoder-only fine-tune of CohereLabs/cohere-transcribe-arabic-07-2026 for Arabic/English code-switching.

Arabic podcast speech routinely borrows English words, and a plain Arabic ASR model transcribes them phonetically in Arabic script ("الويكند"). This model is trained to emit them in Latin script instead ("ال weekend"), which is what downstream consumers of the transcript actually want.

What was trained

Variant middle — the middle third of the decoder layers were unfrozen. The Conformer audio encoder is frozen in every variant of this series, including full; "full" means the full decoder, not the full model.

Training data

Ahmed1/cohere-asr-cs — VAD-segmented Arabic podcast clips (16 kHz mono), transcribed with the base model and then rewritten so English loanwords appear in Latin script. Splits are grouped by source episode, so no episode appears in more than one split.

Trained on 1,900 clips, validated on 454.

Hyperparameters

Final metrics: {"train_runtime": 408.5423, "train_samples_per_second": 27.904, "train_steps_per_second": 0.881, "total_flos": 5.372639220298678e+19, "train_loss": 0.14237111575073666, "epoch": 6.0}

Evaluation

Code-switching (the task)

Measured on 464 held-out clips, 125 of which contain English loanwords (216 Latin word tokens).

Transcription (context, not the target)

WER rises relative to the base model, and that is expected. The references are the base model's own transcriptions with loanwords rewritten into Latin script, so the base reproduces them almost exactly (~1% WER) while never code-switching. Every script change this model makes correctly still counts as edits against a base-shaped reference.

Usage

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor

processor = AutoProcessor.from_pretrained("nsa01n/cohere-cs-decoder-middle")
model = AutoModelForSpeechSeq2Seq.from_pretrained("nsa01n/cohere-cs-decoder-middle")

Feed it 16 kHz mono audio, ideally VAD-segmented to under 30 s per clip — that is how it was trained and the base model caps at 35 s.

Limitations

• Two podcast shows only (sawalef-business, soqrat); other domains and dialects are out of distribution.
• Training targets were machine-generated (base-model transcription + an LLM rewrite pass), not human-verified, so its ceiling is the base model's accuracy on this audio.
• The encoder was never trained, so acoustic robustness is unchanged from the base model.