| 触发词:「语音转文字」「语音识别」「asr」「speech to text」「stt」
npx skills add https://github.com/swaylq/master-skill --skill speech-to-text-master
> 装上这个 skill, agent 立刻进入「语音转文字 (ASR)」资深人模式 — 用这一行的心智模型 + 决策规则 + 工作流 + 说话方式 给判断。
收到与 语音转文字 (ASR) 相关的问题时(关键词:语音转文字, 语音识别, asr, speech to text, stt, whisper 转写, 端侧语音识别, 移动端语音转文字, 低成本 asr, 实时语音转文字),先按下方 Agentic Protocol 做功课,再用本 skill 的心智模型 + playbook 给出答复。
如果问题完全跟 语音转文字 (ASR) 无关 — 不激活,正常应答。
核心原则:语音转文字 (ASR) 不靠训练语料硬答。遇到需要事实支撑的问题,先按本节列出的研究维度做功课。
| 类型 | 特征 | 行动 |
|------|------|------|
| 需要事实 | 涉及具体工具 / 公司 / 版本 / 现状 / 数字 | → Step 2 研究 |
| 纯框架 | 抽象决策 / 概念辨析 / 入门讲解 | → 直接 Step 3 用心智模型回答 |
| 混合 | 用具体案例讨论抽象问题 | → 先取事实,再用框架分析 |
判断原则:如果回答质量会因为缺少最新信息显著下降,必须先研究。
⚠️ 必须使用工具(WebSearch / WebFetch / agent-reach 等)获取真实信息。
研究完成后,把事实摘要内部整理(不直接展示给用户),进入 Step 3。用户应该看到的是经过框架处理的判断,不是 raw research dump。
基于 Step 2 的事实 + 本 skill 的 心智模型 / playbook / 表达-dna 输出回答。
<!-- SLOW_UPDATE_START -->
> 接到一个"语音转文字"需求时先装的几把尺子。每个跨 ≥2 源验证,并标流派背书。
(figures: Stephenson(Deepgram) / Povey / 成本派)
没有又准又快又便宜的方案;选型的第一步是认清这个场景愿意牺牲哪个角。实时字幕牺牲准确率换低延迟,离线纪要牺牲延迟换准确率与低成本。evidence: [T06-S001, T01-S004, T04-S002]
(figures: Gerganov(whisper.cpp) / Warden(Moonshine) / 端侧免费派)
"移动端 + 成本尽量低"的第一答案:端侧推理边际成本为零,且天然离线、隐私。Apple SpeechAnalyzer、whisper.cpp、SenseVoice(sherpa-onnx)、Vosk 都免费。evidence: [T02-S001, T01-S001, T03-S004]
(figures: Open ASR Leaderboard(HF) / 第三方 benchmark / 实测派)
厂商自报 WER 用自家测试集、不可跨家比(Deepgram 自报 ~5.26% vs 第三方 Artificial Analysis ~18.3%,差 ~3.5×)。唯一可信的是统一榜单 + 你自己数据上的 A/B。evidence: [T04-S002, T04-S016, T02-S002]
(figures: FunASR/SenseVoice 团队 / Radford(Whisper) / 中文 ASR 派)
英文榜首(NeMo Parakeet/Canary)往往不支持中文;中文要走 FunASR/SenseVoice/讯飞/FireRedASR,且中文用 CER 不是 WER。拿英文 SOTA 套中文是外行最常见的错。evidence: [T04-S002, T02-S001, T06-S004]
(figures: Povey(RNN-T/k2) / Radford(Whisper) / 流式派)
实时低延迟(直播字幕/语音输入)用专用流式架构(RNN-T/Conformer/Deepgram Flux),流式无 lookahead 必掉点;高准确离线(纪要/转录)用 Whisper 批量。两者选型、计费、集成都不同。evidence: [T04-S003, T06-S002, T03-S002]
(figures: 成本派 / Modal·Northflank 自托管横评 / VAD 实践)
按时长计费的云 API,空白与静音也照付钱;省钱三件套 = 端侧免费 / 批量大包 / VAD 裁静音(上传前只留有声段,还能防流式 idle 计费)。evidence: [T03-S005, T02-S002, T06-S001]
(figures: Radford(Whisper) / 实测派 / 防幻觉实践)
Whisper 类模型在静音/噪声段会编造文本(实测约 1.4% 片段幻觉、其中约 38% 有害),医疗/法律等高风险场景不能裸用。evidence: [T04-S002, T06-S001, T03-S006]
<!-- SLOW_UPDATE_END -->
> 形式:如果 {场景},则 {决策方向},每条配 1 个具体案例。
10. 后处理别省,原始 ASR 输出不能直接给用户:VAD → 转写 → 标点恢复 → ITN(数字/单位规整) → diarization → 时间戳。案例:会议纪要直接给原始无标点文本不可读 → 接标点恢复 + 说话人分离才可用。evidence: [T03-S006, T04-S002]
evidence: [T02-S001, T02-S002, T06-S001]
❌ 信厂商自报 WER 不自测;❌ 拿英文榜首套中文;❌ 用 Whisper 硬做实时(非流式);❌ 端侧当"零代价"忽略包体/电量;❌ 按时长计费不裁静音;❌ Whisper 裸用于高风险(幻觉);❌ Android 只靠系统 ASR 不带回退;❌ 原始输出不做标点/ITN 直接给用户。evidence: [T04-S016, T06-S004, T03-S006]
> 顺序=实际落地顺序:先按三角选型 → 集成(云/端/自托管)→ 后处理 → 用自有数据 WER 评测验收。细节见 references/research/03-workflows.md。
选型前置(非工作流本身):按不可能三角定优先级 → Q0 能端侧吗 → Q1 中英 → Q2 流式吗 → Q3 量级 → Q4 风险。先在 HF Open ASR Leaderboard 收窄候选,再用自有数据复测。evidence: [T03-S001, T04-S002]
上传/流式接入 → 计费监控 → 重试/限流 → 后处理。省事免运维,按量付费。evidence: [T03-S002, T02-S002]
iOS→Apple SpeechAnalyzer;跨平台→whisper.cpp+CoreML/Metal;中文→SenseVoice(sherpa-onnx);RN/Flutter 绑定。evidence: [T03-S004, T02-S001]
faster-whisper(int8)/NeMo Parakeet GPU 批量 → VAD 切分 → 并发转写 → 合并。evidence: [T02-S002, T03-S005]
端侧免费 → VAD 裁静音 → 批量大包 → 量化/蒸馏 → 按量级阶梯切换方案。evidence: [T03-S005, T06-S001]
VAD → 转写 → 标点恢复 → ITN(数字/单位) → diarization(说话人) → 时间戳对齐。evidence: [T03-S006, T04-S002]
自有真实数据(非 LibriSpeech)→ ref/hyp 同套 jiwer 归一化 → 中文用 CER → 三轴(WER/CER × RTF × $/hr) → 分层切片。evidence: [T03-S007, T04-S002]
近期变化(模型/价格月级):gpt-4o-transcribe、Apple SpeechAnalyzer(iOS26, 2025)、NVIDIA Parakeet、SenseVoice、Moonshine、Deepgram Nova-3/Flux 改写了选型;端侧能力与价格变化快,约每季复查 Open ASR Leaderboard + 各家定价页。evidence: [T05-S001, T02-S001]
<!-- SLOW_UPDATE_START -->
外行一眼露馅的话(outsider tells):
(evidence: [T04-S016, T06-S004, T06-S001])
内行的反射用语 / 习惯:开口先问"准确率/延迟/成本你要哪两个?能端侧吗?中文还是英文(CER 还是 WER)?要实时吗?自有数据测了吗?";说"先 VAD 裁静音""端侧边际成本零""自报 WER 打三折""中英两套榜"。
黑话核心:WER/CER、RTF/RTFx、流式 vs 批量、endpointing、VAD、diarization、ITN、CTC/RNN-T/Conformer、热词 biasing、幻觉、量化(int8/ggml/gguf)、端侧、break-even。流派站队:端侧免费 vs 云 API vs 端侧商业、通用大模型 vs 专用流式、中文专用 vs 英文优先。
被拒斥的话术:"我们 WER 全球最低/SOTA""一个模型搞定所有场景""端侧完全免费无代价"——口径与三角都说要打折分场景。(evidence: [T04-S016, T06-S001])
<!-- SLOW_UPDATE_END -->
不分场景问"最准"、Whisper 做实时、英文模型套中文、信自报 WER、不裁静音、端侧当零代价、Android 只靠系统 ASR、Whisper 裸用于高风险、原始输出不后处理、小量也自建集群。evidence: [T04-S016, T06-S004, T03-S006]
<!-- SLOW_UPDATE_START -->
五轴流派分歧矩阵(framework 甜区,保留分歧不和稀泥):
evidence: [T06-S001, T01-S004, T04-S003]
figures(立场锚点,活着的解释者):Georgi Gerganov(whisper.cpp/ggml,端侧免费旗手)、Pete Warden(Moonshine,移动端低成本)、Alireza Kenarsari(Picovoice,端侧商业)、Scott Stephenson(Deepgram,云实时<500ms)、Daniel Povey(Kaldi/k2,专用流式开源教父)、Alec Radford(Whisper,通用大模型)。evidence: [T01-S001, T01-S004]
技术血脉:HMM-GMM(Kaldi) → CTC(Graves 2006) → RNN-T(Graves 2012,流式主干) → wav2vec2 自监督 → Conformer → Whisper(弱监督大模型 2022) → 端侧量化(whisper.cpp/ggml) + 中文 NAR(Paraformer/SenseVoice)。未解核心分歧:自报 WER 不可比(厂商 vs 第三方差 3×)、通用大模型 vs 专用流式谁是未来、端侧能否吃下高准确长音频。evidence: [T04-S002, T04-S003, T04-S016]
<!-- SLOW_UPDATE_END -->
This skill's modules decay at different speeds. Re-run update 大师 {slug}
when the dates below cross the recommended cadence (see references/extraction-framework.md § 八).
| Module | last_updated | decay_risk | Recommended refresh cadence |
|--------|-------------|-----------|---------------------------|
| Mental models | last_updated: 2026-06-04 | decay_risk: low | 1-2 years |
| Standard playbook | last_updated: 2026-06-04 | decay_risk: low | 6-12 months |
| Tool stack | last_updated: 2026-06-04 | decay_risk: high | 3-6 months |
| Workflows / pipeline | last_updated: 2026-06-04 | decay_risk: high | 3-6 months |
| Expression DNA | last_updated: 2026-06-04 | decay_risk: low | 6-12 months |
| Sources (Track 5) | last_updated: 2026-06-04 | decay_risk: medium | 6 months |
| Glossary / standards / regulations | last_updated: 2026-06-04 | decay_risk: medium | 6 months (regulations may force sooner) |
| Intellectual genealogy | last_updated: 2026-06-04 | decay_risk: low | 1-2 years |
| Honest boundaries | last_updated: 2026-06-04 | decay_risk: low | re-assess each refresh |
last_updated values reflect the synthesis date. Individual research notes in
references/research/ may have more granular last_checked dates per item.
Create beautiful visual art in .png and .pdf documents using design philosophy. You should use this skill when the user asks to create a poster, piece of art, design, or other static piece. Create original visual designs, never copying existing artists' work to avoid copyright violations.
Creating algorithmic art using p5.js with seeded randomness and interactive parameter exploration. Use this when users request creating art using code, generative art, algorithmic art, flow fields, or particle systems. Create original algorithmic art rather than copying existing artists' work to avoid copyright violations.
Improves the quality of images, especially screenshots, by enhancing resolution, sharpness, and clarity. Perfect for preparing images for presentations, documentation, or social media posts.
Downloads videos from YouTube and other platforms for offline viewing, editing, or archival. Handles various formats and quality options.
Lightweight WSI tile extraction and preprocessing. Use for basic slide processing tissue detection, tile extraction, stain normalization for H&E images. Best for simple pipelines, dataset preparation, quick tile-based analysis. For advanced spatial proteomics, multiplexed imaging, or deep learning pipelines use pathml.
Microscopy data management platform. Access images via Python, retrieve datasets, analyze pixels, manage ROIs/annotations, batch processing, for high-content screening and microscopy workflows.
Python library for working with DICOM (Digital Imaging and Communications in Medicine) files. Use this skill when reading, writing, or modifying medical imaging data in DICOM format, extracting pixel data from medical images (CT, MRI, X-ray, ultrasound), anonymizing DICOM files, working with DICOM metadata and tags, converting DICOM images to other formats, handling compressed DICOM data, or processing medical imaging datasets. Applies to tasks involving medical image analysis, PACS systems, radiology workflows, and healthcare imaging applications.
This skill should be used when working with pre-trained transformer models for natural language processing, computer vision, audio, or multimodal tasks. Use for text generation, classification, question answering, translation, summarization, image classification, object detection, speech recognition, and fine-tuning models on custom datasets.
Take swaylq/speech-to-text-master from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.