Models & open-source credits
These tools run locally in your browser. Model files and runtime code are downloaded from Hugging Face and jsDelivr only after you start. Your recordings, transcripts, subtitle edits and speaker embeddings are not uploaded. No paid inference service or API key is required.
这些工具在浏览器本地运行，点击开始后才从 Hugging Face 与 jsDelivr 下载模型和运行环境。音频、转写、字幕修改与说话人嵌入不会上传，无需付费推理服务或 API Key。
Vocal isolation / 人声提取
UVR-MDX-NET-Voc_FT — Ultimate Vocal Remover / MDX-Net model, distributed in Blane187/all_public_uvr_models. The distributor declares MIT. We use the original 66,762,490-byte ONNX file without modifying its weights. Upstream UVR project.
The local adapter uses web-audio-separation 0.3.1, MIT, based on python-audio-separator. Its PFFFT FFT component carries the license notices embedded in the bundled worker. Library license.
Speech to text & subtitles / 语音转文字与字幕
Xenova/whisper-base, multilingual int8 ONNX export of OpenAI Whisper Base. Encoder + merged decoder: 76,908,389 bytes (about 77 MB), plus tokenizer/config files and runtime. The ONNX repository declares Apache 2.0; Apache 2.0 license text. The upstream Whisper project uses MIT. Model weights are used unchanged.
Both speech tools use one model cache, CPU inference and 30-second overlapping windows. Audio is decoded at 16 kHz and the higher-energy channel is used for stereo input. Chinese and English transcription are supported; text is not translated. Inputs: 30 MB / 2 minutes. Subtitles use model-generated segment timestamps; missing ends use the next start or the recording end, flagged for review. Manual split timing is estimated from text length. Edits are validated for chronological, non-overlapping positive cue durations before SRT/VTT export. Models can hallucinate words or get timing wrong, especially with noise or silence. Review before publishing.
两个工具共用本地识别模型，支持中英文转写，不翻译内容。字幕初始时间来自模型片段时间戳；缺失结束时间会补为下一段开始或录音结尾，并提示检查。手动拆分的时间为粗略估计。导出前校验时间、重叠与空文字，仍需要回听校对。
Speaker similarity / 说话人相似度
SpeechBrain ECAPA-TDNN, VoxCeleb. Browser ONNX packaging by Pawls2theWall / Equivl, Apache 2.0. License text. We use the unchanged 42,324,080-byte model_fp16.onnx. Weights are stored as float16 and computations use float32. SpeechBrain-compatible STFT, 80-band filterbank and mean normalization are included in the model graph; input is raw 16 kHz mono audio and output is a 192-dimensional embedding. Credit: Mirco Ravanelli and collaborators, SpeechBrain: A General-Purpose Speech Toolkit (2021).
ECAPA-TDNN replaces the earlier WavLM implementation. Cosine scores from different embedding models cannot be compared directly. No same-person threshold has been calibrated here; scores are not identity probabilities or proof of whether audio is cloned. VoxCeleb training data is mostly English interviews; performance may vary across languages, accents and recording conditions.
ECAPA-TDNN 已替代先前的 WavLM。不同模型的分数不能直接比较；本工具未校准同一人阈值，不提供身份概率，也不能判断声音是否被克隆。模型在不同语言、口音和录音条件下的效果需要持续评测。
Runtime / 推理环境
Transformers.js 4.3.1, Apache 2.0. License. ONNX Runtime Web, MIT: vocal and ECAPA speaker workers 1.30.0; Whisper speech worker 1.31.0-dev.20260914-8d85527a0 (the pinned Transformers.js dependency). License.
Models are pinned to immutable repository revisions, checked against their expected size and SHA-256 on initial download, and stored in Audiomni’s model cache where supported. Cache access can be cleared from the AI workspaces. Your browser may evict cached models. No model weights are bundled with the website.
Limits / 使用范围
Vocal isolation: up to 20 MB / 60 seconds, mono or stereo; local CPU inference; stereo 44.1 kHz, 16-bit PCM WAV exports. Speaker comparison: two 3–15-second single-speaker clips; highest-energy input channel, edge silence trimmed; at most the first 10 seconds of the remaining audio is analyzed per clip, resampled to 16 kHz. Noise, overlap, language, microphones and voice processing can change scores. Use samples you have permission to process.