Cohere-AR/cohere-transcribe-ar-sft

Cohere Transcribe Arabic + English for phone calls (SFT, step 10000 of 10,000)

A full supervised fine-tune of CohereLabs/cohere-transcribe-arabic-07-2026 for phone calls and noisy audio in Arabic (dialects), English and code-switched Arabic-English speech, including narrowband (8 kHz) calls. A community fine-tune: not made, affiliated with or endorsed by Cohere.

This is the latest snapshot, step 10000 (also branch step-10000). Every snapshot of the run is a branch step-<N> (every 2,000 steps); main holds the latest one.

Training

• Supervised fine-tuning of the whole model (Conformer encoder and decoder; the encoder's BatchNorm statistics kept frozen), 10,000 steps of about 0.86 h of audio each on 4 GPUs, bf16 autocast, AdamW with a warmup-stable-decay schedule.
• Data: a mix of 69 sources: Arabic dialect speech, Arabic and English phone calls, English meetings and call-centre conversations, accented English, code-switched Arabic-English, and open-TTS sets for numbers, e-mail addresses and places. Augmentation: telephone channels (codecs, narrowband), babble and competing talkers, coloured noise, hum, simulated rooms, speed and SpecAugment.
• Written conventions of the targets: numbers as digits, English words in Latin script (also inside Arabic speech).

Usage

Greedy decoding, 16 kHz mono input, clips of up to ~30 s (cut longer audio at pauses into chunks of at most 25 s). The model writes its own language header when the decoder prompt stops after <|emo:undefined|>:

import torch
from transformers import AutoProcessor, CohereAsrForConditionalGeneration

repo, rev = "Cohere-AR/cohere-transcribe-ar-sft", "main"
processor = AutoProcessor.from_pretrained(repo, revision=rev)
model = CohereAsrForConditionalGeneration.from_pretrained(repo, revision=rev, dtype=torch.bfloat16).to("cuda").eval()

inputs = processor(audio, sampling_rate=16000, return_tensors="pt", language="ar")   # audio: float32, 16 kHz mono
inputs["decoder_input_ids"] = inputs["decoder_input_ids"][:, :4]                      # automatic language
inputs = inputs.to("cuda")
inputs["input_features"] = inputs["input_features"].to(torch.bfloat16)
out = model.generate(**inputs, max_new_tokens=448, do_sample=False)
text = processor.batch_decode(out[:, 4:], skip_special_tokens=True)[0]

Keep the full prompt (language="ar" or "en") to force a language instead. Requires transformers >= 5.4.

Licences

The base model is Apache-2.0. Some training sets are licensed for non-commercial use only (for example SpokenWOZ and L2-ARCTIC) or share-alike (Earnings-22): check them before any commercial use.