diff --git a/README.md b/README.md index 4ce6029..6e527af 100755 --- a/README.md +++ b/README.md @@ -1483,6 +1483,39 @@ B: 太神奇了！那我们快去生成试试吧！ - **自动归档**: 支持将拼接结果自动保存为 `.txt` 文件，文件名支持**自定义前缀** (save_prefix)，方便回溯。 - **输出**: `STRING` (拼接后的文本)。 +#### 6.2 JSON 提取器 (AIIA JSON Extractor 🔑) + +- **用途**: 从 `STRING` 类型的 JSON 字符串中，按 key 路径提取指定字段的值。无需依赖第三方 JSON 节点，完全基于 `STRING` 类型，兼容所有上下游节点。 +- **功能**: + - **嵌套路径**: 支持 dot 路径 + 数组索引，例如 `data.items[0].name`、`[2].speaker`。 + - **多类型输出**: 同时输出 `STRING`、`INT`、`FLOAT`、`BOOLEAN`，自动安全转换。 + - **鲁棒容错**: JSON 格式错误、key 不存在、类型不匹配等情况**不会崩溃**，统一返回 `fallback` 默认值。 + - **自动清理**: 自动处理 BOM、首尾空白等常见 JSON 字符串问题。 +- **输入**: + - `json_string` (STRING, forceInput) — 待解析的 JSON 字符串 + - `key_path` (STRING) — 提取路径，例: `name`, `data.items[0].text`, `[2].speaker` + - `fallback` (STRING, 可选) — 解析失败时的默认值 +- **输出**: `value` (STRING), `value_int` (INT), `value_float` (FLOAT), `found` (BOOLEAN) + +**路径语法示例:** + +| 路径 | JSON 示例 | 提取结果 | +|---|---|---| +| `name` | `{"name": "Alice"}` | `Alice` | +| `data.count` | `{"data": {"count": 42}}` | `42` | +| `items[0].text` | `{"items": [{"text": "Hello"}]}` | `Hello` | +| `[2].speaker` | `[{}, {}, {"speaker": "B"}]` | `B` | +| *(空)* | `{"a": 1}` | 返回整个 JSON | + +#### 6.3 JSON 构建器 (AIIA JSON Builder 🏗️) + +- **用途**: 将多组 key-value 组装为 JSON 对象字符串，方便下游节点消费。 +- **功能**: + - 支持最多 **4 对** key-value 输入。 + - 值如果是合法 JSON（数组、对象、数字、布尔），会自动解析为对应类型而非字符串。 +- **输入**: `key_1` + `value_1` (必填), `key_2`~`key_4` + `value_2`~`value_4` (可选) +- **输出**: `json_string` (STRING) + --- ## ❓ 故障排查 @@ -1500,6 +1533,18 @@ B: 太神奇了！那我们快去生成试试吧！ ## Changelog +### [1.14.2] - 2026-02-17 + +- **NeMo Diarization 兼容性修复**: 修复 PyTorch 2.10+ 环境下 NeMo 说话人分割(`diarize()`)因 lhotse 1.32 不兼容报 `object.__init__() takes exactly one argument` 的崩溃。 + - 自动检测 PyTorch 版本，仅在 >= 2.10 时 monkey-patch `CutSampler.__init__`，移除已废弃的 `data_source` 参数。 +- **JSON Extractor 节点** (New): 从 `STRING` 类型的 JSON 字符串中按 key 路径提取值，支持嵌套路径和数组索引（如 `data.items[0].name`）。 + - 多类型输出：`STRING`、`INT`、`FLOAT`、`BOOLEAN`。 + - 鲁棒容错：JSON 格式错误、key 不存在时返回 `fallback` 默认值，不会崩溃。 +- **JSON Builder 节点** (New): 将多组 key-value 组装为 JSON 对象字符串，支持最多 4 对输入。 +- **Qwen3-TTS Voice Preset**: 新增 `voice_preset` 下拉框（Female_HQ/Male_HQ/Female/Male），无需手动接入参考音频即可使用内置音色，自动启用 Zero-Shot 模式。 +- **Qwen3-TTS 鲁棒性**: `reference_text` 为空或字面值 `"None"` 时自动切换到 Zero-Shot 模式，防止意外走入慢速 ICL 模式。 +- **CosyVoice 参考音频限制**: 参考音频超过 30 秒时自动截断，防止 `AssertionError`。 + ### [1.12.0] - 2026-02-16 - **Podcast Stitcher - MMS Forced Alignment 集成**: 新增 `use_forced_align` 开关，使用 Facebook MMS 声学模型（~1.2GB）进行字级强制对齐。 diff --git a/__init__.py b/__init__.py index e3b6718..7733cac 100755 --- a/__init__.py +++ b/__init__.py @@ -154,6 +154,9 @@ else: # 31. 处理 aiia_text_splitter.py (单人文本拆分) _load_nodes_from_module(".aiia_text_splitter", "aiia_text_splitter") + # 32. 处理 aiia_json_extractor.py (JSON 提取/构建) + _load_nodes_from_module(".aiia_json_extractor", "aiia_json_extractor") + # 告诉 ComfyUI 这个节点包有一个包含网页资源的 'js' 目录 WEB_DIRECTORY = "js" diff --git a/aiia_cosyvoice_nodes.py b/aiia_cosyvoice_nodes.py index b466749..87f2e68 100755 --- a/aiia_cosyvoice_nodes.py +++ b/aiia_cosyvoice_nodes.py @@ -836,6 +836,12 @@ class AIIA_CosyVoice_TTS: ref_wav = torchaudio.transforms.Resample(reference_audio["sample_rate"], sample_rate)(ref_wav) if ref_wav.abs().max() > 1.0: ref_wav = ref_wav / ref_wav.abs().max() + + # Auto-truncate to 30s (CosyVoice _extract_speech_token limit) + max_ref_samples = sample_rate * 30 + if ref_wav.shape[-1] > max_ref_samples: + print(f"[AIIA] Reference audio is {ref_wav.shape[-1]/sample_rate:.1f}s, auto-truncating to 30s.") + ref_wav = ref_wav[..., :max_ref_samples] with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_ref: ref_path = tmp_ref.name diff --git a/aiia_generate_segments.py b/aiia_generate_segments.py index aa99956..c2cd04c 100755 --- a/aiia_generate_segments.py +++ b/aiia_generate_segments.py @@ -7,6 +7,47 @@ from omegaconf import OmegaConf, open_dict # open_dict 允许添加新键 import folder_paths # ComfyUI 的路径管理模块 from typing import Optional, List, Dict, Union # 确保导入类型提示 +# --- Fix: lhotse 1.32 / PyTorch 2.10+ 兼容性补丁 --- +# PyTorch 2.10 移除了 Sampler.__init__(data_source=...) 参数， +# 但 lhotse 的 CutSampler 仍然传递它，导致: +# TypeError: object.__init__() takes exactly one argument +# 在这里进行一次性修复，使 NeMo diarization 能正常工作。 +try: + from torch.utils.data import Sampler as _TorchSampler + # 检测 Sampler 是否已经不接受 data_source (PyTorch >= 2.10) + import inspect + _sampler_params = inspect.signature(_TorchSampler.__init__).parameters + if 'data_source' not in _sampler_params: + try: + from lhotse.dataset.sampling.base import CutSampler as _LhotseCutSampler + _original_init = _LhotseCutSampler.__init__ + + def _patched_cut_sampler_init(self, *args, **kwargs): + # 绕过 super().__init__(data_source=None)，直接初始化 Sampler + _TorchSampler.__init__(self) + # 执行 CutSampler 自身的初始化逻辑 + self.drop_last = kwargs.get('drop_last', False) + self.shuffle = kwargs.get('shuffle', False) + self.seed = kwargs.get('seed', 0) + self.epoch = 0 + from lhotse.dataset.sampling.base import SamplingDiagnostics, _filter_nothing + self._diagnostics = SamplingDiagnostics() + self._just_restored_state = False + self._maybe_init_distributed( + world_size=kwargs.get('world_size', None), + rank=kwargs.get('rank', None) + ) + from lhotse.cut import Cut + self._filter_fn = _filter_nothing() + self._transforms = [] + + _LhotseCutSampler.__init__ = _patched_cut_sampler_init + print("[AIIA] ✅ 已修补 lhotse CutSampler 以兼容 PyTorch 2.10+") + except ImportError: + pass # lhotse 未安装，无需修补 +except Exception as _patch_err: + print(f"[AIIA] ⚠️ lhotse 兼容性补丁失败: {_patch_err}") + # --- 全局模型路径定义 --- _NEMO_MODELS_SUBDIR_STR = "nemo_models" _E2E_MODEL_FILENAME_MAP = { # 支持多种E2E模型 diff --git a/aiia_json_extractor.py b/aiia_json_extractor.py new file mode 100644 index 0000000..3fee22e --- /dev/null +++ b/aiia_json_extractor.py @@ -0,0 +1,208 @@ +""" +AIIA JSON Extractor — 从 STRING 中提取 JSON 某个 key 的值 + +支持嵌套 key 路径 (e.g. "data.items[0].name")， +异常输入时返回 fallback 值而非崩溃。 +""" +import json +import re + +class AIIA_JSON_Extractor: + """从 JSON 字符串中按 key 路径提取值，返回 STRING。""" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "json_string": ("STRING", {"forceInput": True}), + "key_path": ("STRING", { + "default": "", + "tooltip": "提取路径，支持嵌套和数组索引。例: name, data.items[0].text, [2].speaker" + }), + }, + "optional": { + "fallback": ("STRING", { + "default": "", + "tooltip": "解析失败或 key 不存在时返回的默认值" + }), + } + } + + RETURN_TYPES = ("STRING", "INT", "FLOAT", "BOOLEAN",) + RETURN_NAMES = ("value", "value_int", "value_float", "found",) + FUNCTION = "extract" + CATEGORY = "AIIA/Utils" + + @staticmethod + def _parse_path(key_path): + """ + Parse a dot-separated key path with optional array indices. + Examples: + "name" -> ["name"] + "data.items[0]" -> ["data", "items", 0] + "[2].speaker" -> [2, "speaker"] + "lines[0].text" -> ["lines", 0, "text"] + """ + if not key_path or not key_path.strip(): + return [] + + tokens = [] + # Split by '.' first, then handle [N] inside each part + for part in key_path.strip().split('.'): + if not part: + continue + # Match segments like "items[0]" or "[3]" + sub_parts = re.findall(r'([^\[\]]+)|\[(\d+)\]', part) + for name, idx in sub_parts: + if idx != '': + tokens.append(int(idx)) + elif name: + tokens.append(name) + return tokens + + @staticmethod + def _navigate(data, tokens): + """Walk into data following tokens. Returns (value, found).""" + current = data + for token in tokens: + try: + if isinstance(token, int): + if isinstance(current, (list, tuple)) and -len(current) <= token < len(current): + current = current[token] + else: + return None, False + elif isinstance(current, dict): + if token in current: + current = current[token] + else: + return None, False + else: + return None, False + except (KeyError, IndexError, TypeError): + return None, False + return current, True + + def extract(self, json_string, key_path, fallback=""): + log = "[AIIA JSON Extractor]" + + # --- 1. Parse JSON --- + data = None + try: + data = json.loads(json_string) + except (json.JSONDecodeError, TypeError) as e: + print(f"{log} JSON 解析失败: {e}") + # Try to recover: strip leading/trailing whitespace, BOM, etc. + if isinstance(json_string, str): + cleaned = json_string.strip().lstrip('\ufeff') + try: + data = json.loads(cleaned) + print(f"{log} 清理后解析成功") + except Exception: + pass + + if data is None: + print(f"{log} 无法解析 JSON，返回 fallback: {fallback[:50]}") + return (fallback, 0, 0.0, False) + + # --- 2. Navigate key path --- + tokens = self._parse_path(key_path) + + if not tokens: + # No key path: return the entire JSON re-serialized + if isinstance(data, (dict, list)): + result_str = json.dumps(data, ensure_ascii=False, indent=2) + else: + result_str = str(data) + return (result_str, self._safe_int(data), self._safe_float(data), True) + + value, found = self._navigate(data, tokens) + + if not found: + print(f"{log} key '{key_path}' 未找到，返回 fallback") + return (fallback, 0, 0.0, False) + + # --- 3. Convert to output types --- + if isinstance(value, (dict, list)): + value_str = json.dumps(value, ensure_ascii=False, indent=2) + elif value is None: + value_str = "" + elif isinstance(value, bool): + value_str = str(value).lower() + else: + value_str = str(value) + + return (value_str, self._safe_int(value), self._safe_float(value), True) + + @staticmethod + def _safe_int(value): + try: + if isinstance(value, bool): + return 1 if value else 0 + return int(value) + except (ValueError, TypeError): + return 0 + + @staticmethod + def _safe_float(value): + try: + if isinstance(value, bool): + return 1.0 if value else 0.0 + return float(value) + except (ValueError, TypeError): + return 0.0 + + +class AIIA_JSON_Builder: + """将多个 STRING 输入组装为 JSON 对象字符串。""" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "key_1": ("STRING", {"default": "key1"}), + "value_1": ("STRING", {"default": "", "forceInput": True}), + }, + "optional": { + "key_2": ("STRING", {"default": ""}), + "value_2": ("STRING", {"default": "", "forceInput": True}), + "key_3": ("STRING", {"default": ""}), + "value_3": ("STRING", {"default": "", "forceInput": True}), + "key_4": ("STRING", {"default": ""}), + "value_4": ("STRING", {"default": "", "forceInput": True}), + } + } + + RETURN_TYPES = ("STRING",) + RETURN_NAMES = ("json_string",) + FUNCTION = "build" + CATEGORY = "AIIA/Utils" + + def build(self, key_1, value_1, **kwargs): + result = {} + pairs = [(key_1, value_1)] + for i in range(2, 5): + k = kwargs.get(f"key_{i}", "") + v = kwargs.get(f"value_{i}", "") + if k and k.strip(): + pairs.append((k.strip(), v)) + + for key, val in pairs: + # Try to parse value as JSON (for nested objects/arrays/numbers/bools) + try: + parsed = json.loads(val) + result[key] = parsed + except (json.JSONDecodeError, TypeError): + result[key] = val + + return (json.dumps(result, ensure_ascii=False, indent=2),) + + +NODE_CLASS_MAPPINGS = { + "AIIA_JSON_Extractor": AIIA_JSON_Extractor, + "AIIA_JSON_Builder": AIIA_JSON_Builder, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "AIIA_JSON_Extractor": "AIIA JSON Extractor 🔑", + "AIIA_JSON_Builder": "AIIA JSON Builder 🏗️", +} diff --git a/aiia_qwen_nodes.py b/aiia_qwen_nodes.py index 59763ac..a43ef13 100755 --- a/aiia_qwen_nodes.py +++ b/aiia_qwen_nodes.py @@ -262,6 +262,7 @@ class AIIA_Qwen_TTS: }, "optional": { "speaker": (QWEN_SPEAKER_LIST, {"default": "Vivian"}), + "voice_preset": (["None", "Female_HQ", "Male_HQ", "Female", "Male"], {"default": "None", "tooltip": "Built-in voice preset for Base/Clone models. Used when no reference_audio is connected. Auto-enables zero-shot mode."}), "instruct": ("STRING", {"multiline": True, "default": ""}), "reference_audio": ("AUDIO",), "reference_text": ("STRING", {"multiline": True, "default": ""}), @@ -290,6 +291,7 @@ class AIIA_Qwen_TTS: reference_audio = kwargs.get("reference_audio", None) reference_text = kwargs.get("reference_text", "") zero_shot_mode = kwargs.get("zero_shot_mode", False) + voice_preset = kwargs.get("voice_preset", "None") emotion = kwargs.get("emotion", "None") dialect = kwargs.get("dialect", "None") seed = kwargs.get("seed", 42) @@ -299,6 +301,27 @@ class AIIA_Qwen_TTS: top_k = kwargs.get("top_k", 20) top_p = kwargs.get("top_p", 0.95) + # --- Voice Preset Fallback: auto-load seed audio when no reference_audio --- + if reference_audio is None and voice_preset != "None": + import torchaudio + preset_map = { + "Female_HQ": "seed_female_hq.wav", + "Male_HQ": "seed_male_hq.wav", + "Female": "seed_female.wav", + "Male": "seed_male.wav", + } + assets_dir = os.path.join(os.path.dirname(os.path.abspath(__file__)), "assets") + preset_path = os.path.join(assets_dir, preset_map.get(voice_preset, "seed_female_hq.wav")) + if os.path.exists(preset_path): + wav, sr = torchaudio.load(preset_path) + if wav.shape[0] > 1: + wav = torch.mean(wav, dim=0, keepdim=True) + reference_audio = {"waveform": wav.unsqueeze(0), "sample_rate": sr} + zero_shot_mode = True # Preset seeds have no matching text + print(f"[AIIA] Qwen3-TTS: Using voice preset '{voice_preset}' as reference audio (zero-shot mode).") + else: + print(f"[AIIA Warning] Voice preset file not found: {preset_path}") + # 0. Handle Bundle Routing active_qwen = qwen_model if qwen_model.get("is_bundle"): @@ -485,7 +508,7 @@ class AIIA_Qwen_TTS: ref_audio_data = (ref_wav.squeeze().cpu().numpy(), ref_sr) - ref_text = reference_text if reference_text and reference_text.strip() != "" else None + ref_text = reference_text if reference_text and reference_text.strip() not in ("", "None", "none") else None mode_param = zero_shot_mode # Robustness: Qwen requires ref_text for ICL mode (zero_shot_mode=False) diff --git a/pyproject.toml b/pyproject.toml index 7233ddc..3064771 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,7 +1,7 @@ [project] name = "aiia" description = "Advanced AI Audio/Video toolkit for ComfyUI. Features Multi-Role Podcast/Dialogue Generation, High-Fidelity Voice Cloning (CosyVoice/VibeVoice), TTS, Media Management, and efficient Video tools." -version = "1.14.1" +version = "1.14.2" license = {file = "LICENSE"} readme = "README.md" authors = [