dataocean-dolphin-asr

Model Card for ppu-dolphin-asr-onnx

<!-- Provide a quick summary of what the model is/does. -->

This repository provides a collection of ONNX and ORT models for Dolphin, a state-of-the-art multilingual and multitask Automatic Speech Recognition (ASR) model. It features optimizations for various hardware architectures and precision levels.

This modelcard aims to be a base template for new models. It has been generated using this raw template.

Model Details

Model Description

Dolphin is a state-of-the-art multilingual, multitask ASR model developed through a collaboration between Dataocean AI and Tsinghua University. This specific repository (ppu-dolphin-asr-onnx) contains exported ONNX variants of the model.

Key Features:

• 40 Eastern Languages: Supports languages across East Asia, South Asia, Southeast Asia, and the Middle East.

• 22 Chinese Dialects: Comprehensive Chinese dialect support.

• Multiple Variants: Available in standard (dolphin-base) and lightweight (dolphin-small) sizes.

• Quantization Options: Available in No Quantization (Full Precision), FP16 (Half Precision), and INT8 (8-bit quantization).

• Developed by: DataoceanAI and Tsinghua University

• Funded by [optional]: DataoceanAI and Tsinghua University

• Shared by [optional]: Creator of the ppu-dolphin-asr-onnx repository. Conversion scripts provided by DakeQQ.

• Model type: Automatic Speech Recognition (ASR) / ONNX Model

• Language(s) (NLP): 40 Eastern Languages and 22 Chinese Dialects

• License: Please refer to the original Dolphin repository for licensing information.

• Finetuned from model [optional]: Base Dolphin models.

Model Sources

• Repository:
  • Original Repo: DataoceanAI/Dolphin
  • ONNX Conversion scripts: DakeQQ/Automatic-Speech-Recognition-ASR-ONNX
• Paper: arXiv:2503.20212
• Demo: Google Colab Notebook

Uses

Direct Use

The models can be used for Automatic Speech Recognition tasks directly via ONNX Runtime on supported architectures (AMD64 Desktop and ARM processors). The dolphin-small INT8 variant is specifically suited for fast inference in resource-constrained environments.

Downstream Use

This model collection is explicitly designed for integration with the upcoming ppu-voiceland package.

Out-of-Scope Use

Information not provided in the source text.

Bias, Risks, and Limitations

Information not provided in the source text.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

How to Get Started with the Model

You can try out the model and learn how to export your own versions using the provided Google Colab Notebook.

For ecosystem use, integrate this model collection with the ppu-voiceland package.

Training Details

Training Data

The base Dolphin model was trained on 210,000+ hours of extensive proprietary and open-source datasets.

Training Procedure

Preprocessing

All models share the same tokenizer, utilizing the provided tokens.txt file.

Training Hyperparameters

• Training regime: Not specified for original training, but for the ONNX export phase, all models were exported utilizing Greedy Search (not Beam Search).
• Precision: Models are available in Full Precision (No Quantization), FP16, and INT8 formats.

Speeds, Sizes, Times

• ** dolphin-base:** Standard base model with full precision (highest accuracy).
• ** dolphin-small:** Lightweight variant for resource-constrained environments.
• ** Formats:** Available in both .onnx and .ort formats.

Evaluation

Testing Data, Factors & Metrics

Information not provided in the source text.

Results

Information not provided in the source text.

Model Examination

Information not provided in the source text.

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: Information not provided in the source text.
• Hours used: Information not provided in the source text.
• Cloud Provider: Information not provided in the source text.
• Compute Region: Information not provided in the source text.
• Carbon Emitted: Information not provided in the source text.

Technical Specifications

Model Architecture and Objective

Dolphin is a multilingual, multitask Automatic Speech Recognition (ASR) model.

Compute Infrastructure

Hardware

The exported ONNX models feature architecture support for:

• AMD64 Desktop
• ARM processors

Software

Designed to run using ONNX Runtime (supports .onnx and .ort extensions).

Citation

BibTeX:

@misc{dolphin2025,
      title={Dolphin ASR}, 
      author={DataoceanAI and Tsinghua University},
      year={2025},
      eprint={2503.20212},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}