abdelmoez98/cohere-transcribe-arabic-07-2026-ONNX

cohere-transcribe-arabic-07-2026 (ONNX, sherpa-onnx format)

ONNX export of CohereLabs/cohere-transcribe-arabic-07-2026 for offline inference with sherpa-onnx (OfflineCohereTranscribeModelConfig).

Files

The onnx/ folder contains the exported graphs. Each variant has an encoder + decoder + shared tokens.txt:

Variant	Encoder	Decoder	Notes
q4f16	encoder.q4f16.onnx + .onnx_data	decoder.q4f16.onnx + .onnx_data	4-bit + fp16 (~1.5 GB)
int8	encoder.int8.onnx + .onnx_data	decoder.int8.onnx + .onnx_data	dynamic int8 (~3 GB)

Each *.onnx is the graph; the *.onnx_data sidecar holds the weights (external data format, since a 2B-param graph exceeds ONNX's 2 GB protobuf limit).

sherpa-onnx I/O contract

• Encoder: 1 input (mel features) -> 2 outputs (cross_k, cross_v), each [8, batch, 8, seq, 128]. Runs once per utterance.
• Decoder: 6 inputs (tokens, self_k, self_v, cross_k, cross_v, offset) -> 3 outputs (logits, self_k_out, self_v_out).
• The encoder precomputes the cross-attention K/V; the decoder reuses them statically and only maintains its own self-attention cache.
• model_type metadata = cohere-transcribe.

Usage (sherpa-onnx, Python)

import sherpa_onnx
recognizer = sherpa_onnx.OfflineRecognizer.from_cohere_transcribe(
    encoder="./onnx/encoder.q4f16.onnx",
    decoder="./onnx/decoder.q4f16.onnx",
    tokens="./onnx/tokens.txt",
    debug=False,
)
stream = recognizer.create_stream()
stream.accept_waveform(16000, audio_samples)
stream.set_option("language", "ar")
recognizer.decode_stream(stream)
print(stream.result.text)

How this was produced

Exported with tools/cohere-conversion/convert_direct.py (direct torch.onnx.export, no optimum) from transformers 5.14.1 + torch 2.10.0. Run convert_direct.py --variant q4f16 (or int8) then upload_to_hub.py.

License

Derived from the gated CohereLabs/cohere-transcribe-arabic-07-2026. See that model's license terms.