Pitch检测 + NMS，逻辑与demo_pitch_nms.py完全一致

Analyze audio structure and return Level 1 sections. Results are cached based on audio file path, mtime, and parameters. Arg

Main function to analyze subtitles and identify characters. Args: srt_path: Path to the SRT subtitle file movie_name: Name o

Return a JSON schema for the given function.

Parse external SRT and assign speakers via diarization only (no transcription). Args: srt_path: Path to existing SRT file. a

Build a single intro clip from hook_dialogue; timed_lines carry per-line subtitle timing.

Calculate the optimal crop center based on protagonist bounding boxes. VLM bounding boxes are in 1000x1000 coordinate space and are scaled t

Sync wrapper around acall_audio_api. Args: audio_path: Path to audio file prompt: Text prompt for analysis temperatu

Sync wrapper around acall_audio_api_batch. Args: audio_paths: List of paths to audio files prompt: Text prompt for analysis

Check face quality using VLM frame-by-frame detection (same logic as check_face_quality but using VLM). This function loops through

Validate basic structure of the scene proposal (flat format). Returns (passed, feedback_message).

为每个关键点计算综合评分（加权组合多个指标） composite_score = k1 * downbeat_intensity + k2 * pitch_intensity + k3 * mel_energy_intensity 注意：每个指标的强度都已在各自类型内归

计算mel能量 Returns: timestamps: 时间戳数组 energies: (n_frames, n_filters) 的能量矩阵 total_energies: 每帧的总能量 samplerate:

Compute waveform data

Create new subtitle entries with character names instead of speaker labels. Args: subtitles: Original subtitle entries speak

检测音频的Pitch Returns: pitches: pitch数组 confidences: 置信度数组 timestamps: 时间戳数组 actual_samplerate: 实际采样率

Extract all clips from shot_point.json in order. Args: shot_data: List of shot data from shot_point.json cut_points: Optiona

Extract numeric intensity value from string

从镜头数据提取已知人物

Filter keypoints within the specified time range Args: keypoints: List of keypoint dictionaries start_time: Start time in se

Filter keypoints based on sections (Level 1 structure). Args: keypoints: Raw keypoints from Madmom sections:

Collect all sub-segments whose time range overlaps with [start_time_str, end_time_str]. Returns a flat list of sub-segment dicts.

Find the section with highest intensity density (intensity per second), ensuring minimum duration Args: sections: List of sections

Find the section with highest intensity density, or use specified time range Args: audio_caption_path: Path to audio caption JSON fi

找到最近的吸附点

Format speaker dialogues for LLM analysis. Args: speaker_dialogues: Dictionary of speaker to dialogues max_samples: Maximum

Format full subtitles for LLM analysis (preserves conversation flow). Args: subtitles: List of parsed subtitle entries Returns:

Format subtitle list as dialogue lines.

Generate caption JSON from edited keypoints and pre-computed sections. Args: audio_path: Path to audio file

生成场景描述，验证必需字段

Render keypoint markers as a single SVG overlay (fewer DOM nodes than div markers).

Render section regions as colored rectangles on the waveform.

Generate a one-to-one shot mapping for each music segment.

Generate and parse shot plan with validation + retry.

Generate structure proposal with basic validation and retry.

Generate SVG string

Convert audio file to base64 data URL for browser playback

Get audio sample rate using ffprobe.

Get color for keypoint type.

Return operationally important shot-plan parts that are missing.

Return a Pydantic model for the parameters of the given function.

Return a JSON schema for the parameters of the given function.

Return a Pydantic model for the returns of the given function.

Group dialogues by speaker. Args: subtitles: List of parsed subtitle entries Returns: Dictionary mapping speaker labels

Return subtitle entries that overlap [start, end].

Get video framerate using ffprobe.

Check if torchcodec is available and properly installed.

Load audio using audioread (ffmpeg backend) as fallback. Args: path_or_buffer: File path, URL, or buffer sr: Target samp

Load saved parameters from config file

同步读帧（decord not thread-safe，必须在 event loop 主线程调用）

从 caption JSON 文件加载 sections 信息 Args: caption_path: caption.json 文件路径 Returns: sections 列表，每个元素包含 name, start_t

Example usage of audio caption function with Madmom-based segmentation.

Parse config override arguments in the format --config.PARAM_NAME value Args: unknown_args: List of unknown arguments from argparse

Parse shot plan JSON from LLM output, stripping markdown fences if present.

Parse shot_scenes.txt and return list of scene cut timestamps in seconds. Args: shot_scenes_path: Path to shot_scenes.txt file

Parse a shot_scenes.txt file into a list of (start_frame, end_frame) tuples.

Parse an SRT subtitle file into a list of subtitle entries. Args: srt_path: Path to the SRT file Returns: List of dicti

Print a summary of all clips to be rendered.

处理单个文件，scene_data/frames 可由外部预先提供

处理单个场景，frames 可由外部预先读取以避免阻塞 event loop

Query LLM to identify characters from their dialogues. Args: dialogues_text: Formatted dialogue text movie_name: Optional mo

Refine character mapping and handle edge cases. Args: character_info: Raw character identification results speaker_dialogues

Refresh hook dialogue in an existing shot-plan JSON file and save it in place.

Render video clips using ffmpeg concat demuxer. Args: video_path: Path to source video file clips: List of clip dictionaries

Reset parameters to default values

Review whether the shot selection meets requirements. Args: shot_proposal: Shot selection info - answer:

Check if the proposed time range overlaps with any previously used clips. You MUST call this tool BEFORE calling finish to ensure no duplicat

Review and validate the proposed shot selection before finishing. Validates that exactly ONE shot is provided and its duration matches the ta

High-level ASR: extract audio from video, transcribe, write SRT. Args: video_path: Input video file. output_dir: Directory f

Run a single shot selection loop and return the commit result dict on success.

Save parameters to config file

Convert seconds to HH:MM:SS.s format.

Use LLM to select the best music section, then trim to target duration if needed.