Cohere-AR/cohere-transcribe-ar-rl

Cohere Transcribe Arabic + English for phone calls (weight average)

A community 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: not made, affiliated with or endorsed by Cohere.

This is branch main: a weight average of 0.7 x branch v4c-step-800 (SFT v4 + three MWER stages: human-annotated Levantine phone calls cut with the benchmark's edge padding, digit grouping scored, Arabic-English code-switching) and 0.3 x branch r4-step-800 of this repository, tensors averaged in float32; no further training (also branch soup-v4c70-rl4). main holds the best checkpoint of the repository so far.

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-rl", "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, L2-ARCTIC and some code-switching corpora) or share-alike (Earnings-22): check them before any commercial use.