aha/docs/model-card.md

OCR

PaddleOCR-VL

Options prompt:

"OCR:" | "Table Recognition:" | "Formula Recognition:" | "Chart Recognition:"

PaddleOCR-VL1.5

Options prompt:

"OCR:" | "Table Recognition:" | "Formula Recognition:" | "Chart Recognition:" | "Spotting:" | "Seal Recognition:"

DeepSeekOCR

Main Prompts

1. document: \n<|grounding|>Convert the document to markdown.
2. without layouts: \nFree OCR.

Metadata

• base_size: 512 | 640 | 1024 | 1280
• image_size: 512 | 640 | 1024 | 1280
• crop_mode: false | true

Example:

{"base_size": "640", "image_size": "640", "crop_mode": "false"}

Processing Types

Type	base_size	image_size	crop_mode
Tiny	512	512	false
Small	640	640	false
Base	1024	1024	false
Large	1280	1280	false
Gundam	1024	640	true

DeepSeekOCR2

Main Prompts

1. document: \n<|grounding|>Convert the document to markdown.
2. without layouts: \nFree OCR.

Metadata

• crop_mode: false | true

GLM-OCR

Prompt Limited

Document Parsing prompt：

"Text Recognition:" | "Formula Recognition:" | "Table Recognition:"

Information Extraction:

extract structured information from documents. Prompts must follow a strict JSON schema, For example, to extract personal ID information:

Please output the information in the image in the following JSON format:
{
    "id_number": "",
    "last_name": "",
    "first_name": "",
    "date_of_birth": "",
    "address": {
        "street": "",
        "city": "",
        "state": "",
        "zip_code": ""
    },
    "dates": {
        "issue_date": "",
        "expiration_date": ""
    },
    "sex": ""
}

HunyuanOCR

Application-oriented Prompts

Task	English	Chinese
Spotting	Detect and recognize text in the image, and output the text coordinates in a formatted manner.	检测并识别图片中的文字，将文本坐标格式化输出。
Parsing	1. Identify the formula in the image and represent it using LaTeX format. 2.Parse the table in the image into HTML. 3. Parse the chart in the image; use Mermaid format for flowcharts and Markdown for other charts. 4.Extract all information from the main body of the document image and represent it in markdown format, ignoring headers and footers. Tables should be expressed in HTML format, formulas in the document should be represented using LaTeX format, and the parsing should be organized according to the reading order.	1. 识别图片中的公式，用 LaTeX 格式表示。 2. 把图中的表格解析为 HTML。 3. 解析图中的图表，对于流程图使用 Mermaid 格式表示，其他图表使用 Markdown 格式表示。 4. 提取文档图片中正文的所有信息用 markdown 格式表示，其中页眉、页脚部分忽略，表格用 html 格式表达，文档中公式用 latex 格式表示，按照阅读顺序组织进行解析。
Information Extraction	1. Output the value of Key. 2. Extract the content of the fields: ['key1','key2', ...] from the image and return it in JSON format. 3. Extract the subtitles from the image.	1. 输出 Key 的值。 2. 提取图片中的: ['key1','key2', ...] 的字段内容，并按照 JSON 格式返回。 3. 提取图片中的字幕。
Translation	First extract the text, then translate the text content into English. If it is a document, ignore the header and footer. Formulas should be represented in LaTeX format, and tables should be represented in HTML format.	先提取文字，再将文字内容翻译为英文。若是文档，则其中页眉、页脚忽略。公式用latex格式表示，表格用html格式表示。

Here is the continuation for the English model-card.md file, translating and adapting the content from the Chinese version regarding VoxCPM models.

TTS

VoxCPM (0.5B / 1.5)

Mode Description

• Zero-shot TTS (Default): Generates speech directly without reference audio.
• Voice Cloning: Requires reference audio (audio_url) and its corresponding transcript (prompt_text).
  • Note: For versions 0.5B/1.5, if audio_url is provided, prompt_text must be provided in metadata. Conversely, if there is no audio_url, prompt_text should not be provided.

Metadata Parameters

Parameter	Type	Required	Description
prompt_text	String	Conditional	Required for Cloning. The transcript of the reference audio. Used to extract speaker characteristics.
min_len	Integer	No	Minimum generation length (tokens/steps), default 2.
max_len	Integer	No	Maximum generation length (tokens/steps), default 4096.
inference_timesteps	Integer	No	Number of inference steps. Affects quality and speed. Default 10. Higher values yield better quality but slower speed.
cfg_value	Float	No	Classifier-Free Guidance value. Controls adherence to the prompt. Default 2.0.
retry_badcase_ratio_threshold	Float	No	Threshold for retrying bad cases. Default 6.0.

Example 1: Zero-shot TTS (No Reference)

{
    "model": "OpenBMB/VoxCPM-0.5B",
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "type": "text", 
                    "text": "Hello, this is a test speech."
                }
            ]
        }
    ]
}

Example 2: Voice Cloning (With Reference)

{
    "model": "OpenBMB/VoxCPM-0.5B",
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "type": "audio",
                    "audio_url": {
                        "url": "https://example.com/reference.wav"
                    }
                },              
                {
                    "type": "text", 
                    "text": "This is the target text I want to generate."
                }
            ]
        }
    ],
    "metadata": {
        "prompt_text": "This is the text content corresponding to the reference audio."
    }
}

Response Example

The model returns a JSON object containing Base64-encoded audio data. The audio format is WAV.

{
    "id": "chatcmpl-uuid...",
    "object": "chat.completion",
    "created": 1710000000,
    "model": "OpenBMB/VoxCPM-0.5B",
    "choices": [
        {
            "index": 0,
            "message": {
                "role": "assistant",
                "content": [
                    {
                        "type": "audio",
                        "audio_url": {
                            "url": "data:audio/wav;base64,UklGRiQAAABXQVZFZm10IBAAAAABAAEAQB8AAEAfAAABAAgAZGF0YQAAAAA..."
                        }
                    }
                ]
            },
            "finish_reason": "stop"
        }
    ],
    "usage": null
}

VoxCPM2

VoxCPM2 supports three advanced voice generation modes. Switch between modes by combining content types in messages and parameters in metadata.

1. Voice Design

Generate a completely new voice from natural language descriptions alone (gender, age, tone, emotion, pace, etc.). No reference audio required.

• Input: Text message only.
• Metadata: Optional control_instruction for finer control (e.g., emotion, pace).

Example:

{
    "model": "OpenBMB/VoxCPM2",
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "type": "text", 
                    "text": "Welcome to the future world."
                }
            ]
        }
    ],
    "metadata": {
        "control_instruction": "Young female, energetic" 
    }
}

2. Controllable Cloning

Clone a voice from a short audio clip, with optional style guidance to steer emotion, pace, and expression while preserving timbre.

• Input: Reference audio (audio_url) + Target text.
• Metadata: Optional control_instruction to adjust style (emotion, pace, etc.). prompt_text is NOT required.

Example:

{
    "model": "OpenBMB/VoxCPM2",
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "type": "audio",
                    "audio_url": {
                        "url": "https://example.com/short_clip.wav"
                    }
                },
                {
                    "type": "text", 
                    "text": "This news broadcast needs to be more serious and calm."
                }
            ]
        }
    ],
    "metadata": {
        "control_instruction": "serious, slow pace"
    }
}

3. Ultimate Cloning

Provide reference audio and its verbatim transcript for audio-continuation cloning or high-fidelity cloning. Every vocal nuance is faithfully reproduced.

• Input: Reference audio (audio_url) + Target text.
• Metadata: Must provide prompt_text (accurate transcript of the reference audio).

Example:

{
    "model": "OpenBMB/VoxCPM2",
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "type": "audio",
                    "audio_url": {
                        "url": "https://example.com/reference.wav"
                    }
                },
                {
                    "type": "text", 
                    "text": "This is the new content to be spoken next."
                }
            ]
        }
    ],
    "metadata": {
        "prompt_text": "This is the actual text spoken in the reference audio."
    }
}

General Metadata Parameters (Applicable to all VoxCPM2 modes)

Parameter	Type	Required	Description
prompt_text	String	Ultimate Cloning Only	Verbatim transcript of the reference audio. Required only for Ultimate Cloning mode.
control_instruction	String	No	Natural language instruction to control style for Voice Design or Controllable Cloning (e.g., "angry", "whispering", "fast").
min_len	Integer	No	Minimum generation length, default 2.
max_len	Integer	No	Maximum generation length, default 4096.
inference_timesteps	Integer	No	Inference steps, default 10.
cfg_value	Float	No	CFG value, default 2.0.
retry_badcase_ratio_threshold	Float	No	Bad case retry threshold, default 6.0.

Response Example

Same as VoxCPM 0.5B/1.5, returns a JSON object containing Base64-encoded WAV audio.

{
    "id": "chatcmpl-uuid...",
    "object": "chat.completion",
    "created": 1710000000,
    "model": "OpenBMB/VoxCPM2",
    "choices": [
        {
            "index": 0,
            "message": {
                "role": "assistant",
                "content": [
                    {
                        "type": "audio",
                        "audio_url": {
                            "url": "data:audio/wav;base64,UklGRiQAAABXQVZFZm10IBAAAAABAAEAQB8AAEAfAAABAAgAZGF0YQAAAAA..."
                        }
                    }
                ]
            },
            "finish_reason": "stop"
        }
    ],
    "usage": null
}