MCP Video Extraction Plus

🚀 MCP Video Extraction plus - Integration Modification Instructions for Speech Recognition Service

This project is forked from MCP Video Extraction. This modification expands the project's speech recognition service, originally supporting only Whisper (local model), to three options:

1. Whisper – OpenAI's local speech recognition model (original)
2. JianYing (CapCut) – ByteDance's JianYing/CapCut online speech recognition service
3. Bcut (Bilibili Cut) – Bilibili's online speech recognition service

📦 Newly Added Files

Core ASR Module

📝 Modified Files

1. src/mcp_video_service/services/video_service.py

Major Changes:

• Added support for ASR provider configuration (asr_provider).
• Modified the __init__ method to support reading the ASR provider selection from the configuration file.
• Added a new _create_asr_instance() method to create the corresponding ASR instance based on the configuration.
• Modified the extract_text() method to support multiple ASR providers:
  • If Whisper is selected, use the local model.
  • If JianYing or Bcut is selected, call the online API.
• Added import statements to support both relative and absolute imports.

Configuration Parameters:

self.config = {
    'asr': {
        'provider': 'whisper',  # Options: whisper, jianying, bcut
        'use_cache': False,
        'need_word_time_stamp': False,
    },
    'jianying': {
        'start_time': 0,
        'end_time': 6000,
    },
    # ... Other configurations
}

2. config.yaml

Newly Added Configuration Items:

# ASR (Automatic Speech Recognition) provider configuration
asr:
  provider: "whisper"  # Options: whisper, jianying, bcut
  use_cache: false
  need_word_time_stamp: false

# JianYing (CapCut) speech recognition configuration
jianying:
  start_time: 0      # Audio start time (milliseconds)
  end_time: 6000     # Audio end time (milliseconds)

# Bcut (Bilibili Cut) speech recognition configuration
# No special configuration for now, use default values

💻 Usage Examples

1. Use Whisper (Local Model, Default)

from mcp_video_service.services.video_service import VideoService

service = VideoService(config_path='config.yaml')
# Set in config.yaml: asr.provider = "whisper"

text = await service.extract_text('audio.mp3')

2. Use JianYing (Online Service)

# Modify config.yaml
# asr:
#   provider: "jianying"
#   use_cache: false
#   need_word_time_stamp: false
# jianying:
#   start_time: 0
#   end_time: 6000

service = VideoService(config_path='config.yaml')
text = await service.extract_text('audio.mp3')

3. Use Bcut (Online Service)

# Modify config.yaml
# asr:
#   provider: "bcut"
#   use_cache: true
#   need_word_time_stamp: false

service = VideoService(config_path='config.yaml')
text = await service.extract_text('audio.mp3')

4. Configure via Environment Variables

# Use JianYing
export ASR_PROVIDER=jianying
export ASR_USE_CACHE=false
export JIANYING_START_TIME=0
export JIANYING_END_TIME=6000

# Use Bcut
export ASR_PROVIDER=bcut
export ASR_USE_CACHE=true

🏗️ Architecture Design

Class Inheritance Relationship

BaseASR (Abstract Base Class)
├── JianYingASR
├── BcutASR
└── VideoService (Call whisper.load_model directly when using Whisper)

Key Features of BaseASR

• Unified Interface: All ASR implementations inherit from BaseASR.
• Cache Support: Generate cache keys through CRC32 checksums.
• Rate Limiting: Built - in rate - limiting mechanism to prevent API overload.
• Progress Callback: Support asynchronous progress callback functions.
• Error Handling: Unified error handling and logging.

📊 Data Structures

ASRDataSeg (Segmented Data)

@dataclass
class ASRDataSeg:
    text: str          # Segmented text
    start_time: float  # Start time (milliseconds)
    end_time: float    # End time (milliseconds)

ASRData (Complete Data)

@dataclass
class ASRData:
    text: str                    # Complete text
    segments: List[ASRDataSeg]   # List of segments

📈 Status Management

The status of ASR tasks is managed through the ASRStatus enumeration:

📥 Import Compatibility

All newly added ASR modules support both relative imports and absolute imports:

# Relative import (used within the package)
from .asr_data import ASRDataSeg
from .base_asr import BaseASR

# Absolute import (used when directly loading the module)
from asr_data import ASRDataSeg
from base_asr import BaseASR

This ensures that the modules work properly in different import scenarios.

📦 Dependencies

Newly Added Dependencies

• requests - Used for HTTP requests (JianYing and Bcut APIs)

Existing Dependencies

• yt-dlp - Video download
• whisper - Local speech recognition (only required when using Whisper)
• pyyaml - Configuration file parsing

🌐 Environment Variable Support

⏮️ Backward Compatibility

This modification is fully backward - compatible:

• The default ASR provider is whisper, maintaining the original behavior.
• The original extract_text() interface remains unchanged.
• All new features are optional.

✅ Test Verification

All newly added modules have passed the following verifications:

• ✓ Python syntax check
• ✓ Module import test
• ✓ Class inheritance relationship verification
• ✓ Configuration file parsing test

🔜 Future Expansion

This architecture supports easily adding new ASR providers:

1. Create a new class that inherits from BaseASR.
2. Implement the _run() and _make_segments() methods.
3. Add a conditional branch in VideoService._create_asr_instance().
4. Add corresponding configuration items in config.yaml.

📄 License

The license of the original project remains unchanged.