展示HN：Chirp – 使用ParakeetV3的本地Windows语音输入，无需可执行文件

展示HN：Chirp – 使用ParakeetV3的本地Windows语音输入，无需可执行文件
Show HN: Chirp – Local Windows dictation with ParakeetV3 no executable required

## Chirp：Windows本地快速语音听写 Chirp是一款注重隐私的Windows语音听写应用，完全本地运行，仅需Python即可，无需任何可执行文件。它利用ParakeetV3语音转文本（STT）引擎，提供与OpenAI的Whisper-large-V3相媲美的准确性，但速度**快17倍**，并且仅使用**CPU运行**（Whisper通常需要GPU）。 Chirp具有高度可配置性，允许用户自定义热键（默认：Ctrl+Shift）、语言和后处理样式。它还具有剪贴板管理、音频反馈和自定义词汇覆盖（例如，将“parrakeat”纠正为“parakeet”）等功能。该应用利用ONNX进行高效处理，并提供设置来控制线程使用和模型量化，以优化性能。Chirp通过将所有处理保留在用户机器上，优先考虑用户控制和隐私。

## Chirp：本地Windows语音输入 Chirp是一款新的Windows语音输入应用程序，专为禁止安装可执行文件的受限环境而设计。它由其创建者首次开源项目开发，提供准确、快速且*本地*的语音转文本功能，无需GPU。 Chirp利用NVIDIA的ParakeetV3模型（准确性可与Whisper-large-v3媲美，但在CPU上速度更快），并完全通过Python和`uv`进行管理。它避免使用云服务，而是通过全局热键将识别出的文本直接输入到活动窗口中。用户可以通过`config.toml`文件配置Chirp，调整诸如热键、模型选项和后处理（例如句子大小写）等设置。开发者正在寻求反馈，特别是来自受限环境的用户、熟悉Parakeet/Whisper/ONNX Runtime的用户以及拥有不同硬件/语言设置的用户。一位用户报告说，更喜欢Whisper的自然流畅的转录和自动去除填充词的功能。该项目在GitHub上可用：[https://github.com/Whamp/chirp](https://github.com/Whamp/chirp)。

原文

Chirp is a Windows dictation app that runs fully locally using ParakeetV3 STT and is managed end-to-end with uv. Chirp does not require the ability to run executable files (like .exe) on Windows. It was designed so that if you're allowed to run Python on your machine, you can run Chirp.

ParakeetV3 has indistinguishable accuracy from Whisper-large-V3 (multilingual WER 4.91 vs 5.05) but is 17x faster and only requires a CPU while Whisper models of comparable accuracy require GPU's.

primary_shortcut = "ctrl+shift"                 # Hotkey that toggles recording; any combination supported by the `keyboard` library works (e.g. "ctrl+shift+space").
stt_backend = "parakeet"                        # Only "parakeet" is bundled today, but keeping this key lets us add more backends later if needed.
parakeet_model = "nemo-parakeet-tdt-0.6b-v3"    # Deployed ONNX bundle name; keep as-is unless new models are added.
parakeet_quantization = ""                      # Set to "int8" to download/use the quantized model variant; leave blank for default fp16.
onnx_providers = "cpu"                          # ONNX runtime provider string (comma- or pipe-separated if your build supports multiple providers, e.g. "cuda" or "cpu|dml").
threads = 0                                     # 0 (or empty) lets ONNX decide; set a positive integer to pin thread usage.
language = "en"                                 # Optional ISO language code; leave blank to let Parakeet auto-detect.
post_processing = ""                            # Text prompt for the StyleGuide; see docs/post_processing_style_guide.md (e.g. "sentence case", "prepend: >>", "append: — dictated with Chirp").
paste_mode = "ctrl"                             # Non-Windows platforms honor this: "ctrl" -> Ctrl+V, "ctrl+shift" -> Ctrl+Shift+V. Windows types text directly today.
clipboard_behavior = true                       # Keeps clipboard history clean when true by clearing it after `clipboard_clear_delay` seconds.
clipboard_clear_delay = 0.75                    # Seconds to wait before clearing the clipboard (only if `clipboard_behavior` is true).
audio_feedback = true                           # Enables start/stop chime playback.
start_sound_path = ""                           # Leave blank to use bundled asset; default: src/chirp/assets/ping-up.wav
stop_sound_path = ""                            # Leave blank to use bundled asset; default: src/chirp/assets/ping-down.wav

# Word overrides map spoken tokens (case-insensitive) to replacement text.
[word_overrides]
parrakeat = "parakeet"
"parra keat" = "parakeet"

展示HN：Chirp – 使用ParakeetV3的本地Windows语音输入，无需可执行文件 Show HN: Chirp – Local Windows dictation with ParakeetV3 no executable required

展示HN：Chirp – 使用ParakeetV3的本地Windows语音输入，无需可执行文件
Show HN: Chirp – Local Windows dictation with ParakeetV3 no executable required