FluidInference / cohere-transcribe-03-2026-coreml

Cohere Transcribe 03-2026 - CoreML

CoreML conversion of CohereLabs/cohere-transcribe-03-2026 , a multilingual speech recognition model supporting 14 languages with state-of-the-art accuracy.

Original Model : CohereLabs/cohere-transcribe-03-2026 Converted for : Apple Silicon (Neural Engine) on macOS 14.0+ and iOS 17.0+

About

This repository contains CoreML-optimized versions of the Cohere Transcribe model from Cohere For AI . The original PyTorch model has been converted to CoreML format for efficient on-device inference on Apple platforms using the Neural Engine.

For the original model, documentation, and training details, see: CohereLabs/cohere-transcribe-03-2026

Model Files

This repository contains CoreML models in two variants:

Standard (FP16/FP32)

• cohere_audio_encoder.mlmodelc (~3.6 GB) - Audio feature extraction
• cohere_decoder.mlmodelc (~293 MB) - Transformer decoder with KV cache
• cohere_lm_head.mlmodelc (~32 MB) - Language model head for logits

v2 (INT8 Quantized)

• cohere_audio_encoder_v2.mlmodelc (~1.8 GB) - INT8 quantized audio encoder
• cohere_decoder.mlmodelc (~293 MB) - Same decoder (shared)
• cohere_lm_head.mlmodelc (~32 MB) - Same LM head (shared)

Both .mlmodelc (compiled) and .mlpackage (source) formats are included for each model.

Which version to use?

• Standard : Best accuracy, use when you have sufficient memory (4.5GB)
• v2 (Quantized) : 50% smaller (2.7GB total), minimal accuracy loss, faster on some devices

Supported Languages

English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Arabic, Chinese (Mandarin), Japanese, Korean, Vietnamese

Benchmark Results

Evaluated on FLEURS dataset (100 samples per language):

Overall : 13 languages, 1,300 samples, 3.19x average RTFx

WER = Word Error Rate (Western languages), CER = Character Error Rate (Asian languages)

Usage

FluidAudio Swift Package

import FluidAudio

// Initialize with automatic model download
let manager = try await CohereAsrManager(language: .english)

// Load audio (16kHz mono)
let audioBuffer: AVAudioPCMBuffer = ...

// Transcribe
let result = try await manager.transcribe(audioBuffer)
print(result.text)

Command Line (FluidAudio CLI)

# Transcribe audio file
fluidaudiocli cohere-transcribe audio.wav

# Specify language
fluidaudiocli cohere-transcribe audio.wav --language french

# Benchmark on FLEURS
fluidaudiocli cohere-benchmark --languages en_us,fr_fr --max-files 100

Hardware Requirements

• Minimum : Apple Silicon Mac (M1+) or iPhone with A14+ chip
• Recommended : M2+ for optimal performance
• Memory : ~4.5 GB for all models loaded
• Storage : ~4 GB for compiled models

Architecture

3-model pipeline with streaming decoding:

1. Audio Encoder : Converts mel spectrogram → encoder hidden states (768-dim)
2. Decoder : Autoregressive transformer with KV cache for efficient streaming
3. LM Head : Projects decoder states → vocabulary logits (51,866 tokens)

Special tokens:

• <blank> : 51,865 (CTC blank token)
• Language IDs: 51,851-51,864 (14 languages)

Performance Notes

• RTFx 3.19x : Faster than real-time on Apple Silicon
• Latency : 3-4s average per inference (includes model loading)
• Neural Engine : ~95% of operations run on ANE with default compute units
• Streaming : Supports chunk-based processing with decoder state caching

Integration

Download via ModelRegistry

Models auto-download on first use from HuggingFace:

let models = try await CohereAsrModels.download()
// Returns: CohereAsrModels with paths to encoder, decoder, lm_head

Manual Download

git clone https://huggingface.co/FluidInference/cohere-transcribe-03-2026-coreml

Or use FluidAudio CLI:

fluidaudiocli download --dataset cohere-transcribe

Conversion Details

• Source : PyTorch model from CohereLabs/cohere-transcribe-03-2026
• Tools : coremltools 9.0b1, PyTorch 2.5.1
• Target : iOS 17+, macOS 14+
• Precision : Mixed FP16/FP32 (automatic by coremltools)
• Compute Units : ALL (automatic ANE/GPU/CPU assignment)

License

MIT License (same as original model)

Citation

@misc{cohere-transcribe-03-2026,
  author = {Cohere For AI},
  title = {Cohere Transcribe 03-2026},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/CohereLabs/cohere-transcribe-03-2026}}
}

Links

• Original Model: CohereLabs/cohere-transcribe-03-2026
• FluidAudio Package: github.com/Beingpax/VoiceLink/FluidAudio
• Conversion Pipeline: mobius/models/stt/cohere-transcribe-03-2026