💯 Completely offline. No API keys. No accounts. No cloud. Nothing ever leaves your machine.
|
🚀 Start here |
🔬 What changed |
🛠️ Under the hood |
|
Tap the mic and talk. Your words are typed into whatever window has focus. |
Copy any text, tap Speak Clipboard, hear it in a natural neural voice. |
Every model runs locally. The app never opens a network connection. |
Not one content word correct. No punctuation. No capitals. |
Sentence, contractions, punctuation, capitals. All of it. |
🎧 Same 7-second recording. Same USB microphone. Same moment. Only the engine changed.
Vosk is built on Kaldi, and that generation of recognizer has simply been overtaken. Switching to Vosk's most accurate English model didn't fix it either — it just pushed cold start to 35 seconds.
| 🐌 Before | 🚀 Now | 🎯 Gain | |
|---|---|---|---|
| ⏱️ Time to window | 129.5 s |
2.6 s |
127 seconds faster |
| 🎤 Dictation ready | 129.5 s |
~5 s |
the recognizer loads behind the open window |
| 🔊 First word of a long read | 77 s |
1.2 s |
no longer grows with the text |
| 🧠 Memory | 9.6 GB |
2.3 GB |
7.3 GB lighter |
Note
🕵️ Why the old number was so bad. The old build loaded its 2.7 GB speech model twice — once in the launcher, once in the app, which never read the launcher's copy. That 9.6 GB was two full copies of the same model sitting in RAM.
Note
⏱️ Nothing on screen waits for a model any more. The window used to sit behind the full splash video and a 2.4 GB recognizer load. Speak Clipboard never touches the recognizer, and dictation only needs it to decode — so it loads behind the open window, the microphone opens immediately, and audio is buffered until it arrives. Nothing spoken in those first seconds is lost.
The 77 s above was a 2,500 character paste. It is not a typo: the old reader rendered the
entire clipboard to a WAV before playing a single word, so the silence scaled with the text.
Text-to-speech used to be Piper on en_US-hfc_male-medium. Small, fast, and audibly
synthetic on anything longer than a sentence.
It now runs Kokoro-82M through sherpa-onnx — which was already a dependency, so this added zero new ones.
| 🔈 Piper | 🔊 Kokoro-82M | |
|---|---|---|
| 🧬 Parameters | ~20 M | 82 M |
| 🎭 Naturalness | flat, synthetic | markedly better |
| ⚡ Speed on CPU | fast | ~1.5x realtime |
| 📅 Generation | older | current |
Piper stays available rather than deleted. A few of its voices hold up fine.
python voice_picker.pywClick a voice ▸ hear it read a sample line ▸ Use this voice.
Saved to %APPDATA%/SpeakAnywhere/voice.json, picked up on next launch.
Tip
🧹 The list is short on purpose. Kokoro ships 54 voices and publishes a
quality grade for each.
Everything graded C or D is excluded — am_adam is a D — plus af_sky and am_santa,
both trained on only minutes of audio.
✅ What's left: 15 English voices at grade A or B, plus 3 Piper voices. The point is to pick a voice, not to audition a pile of bad ones.
It starts talking on the first sentence instead of rendering the whole clipboard first.
📋 clipboard
│
▼
sentences ──► chunks: 60 chars ─► 90 ─► 135 ─► 202 ─► 400 …
│
▼
🗣️ render thread ──► 📥 queue (3 deep) ──► 🔈 one output stream
~1.75x realtime gapless
The opener is short because it is the only piece anyone waits on. Everything after it is rendered while earlier audio is still playing.
| Why it is built this way | |
|---|---|
| 📈 | Chunks grow by half each time. Rendering runs ~1.75x faster than speech plays, so every pass banks more finished audio than it spends. Jumping straight to full size leaves the renderer 8 seconds behind after a 2 second opener — a stall exactly where it is most audible. |
| ✂️ | A long opening sentence is broken at its commas. Only near the start, where there is no buffer to coast on. Past that, sentences are left whole so chunk edges land where a reader would pause anyway. |
| 🔗 | One output stream for the whole read, opened on the first chunk's rate — so the joins are silent instead of a click per sentence. |
| ⏹️ | Stop lands within 0.2 s. Audio is written to the device in fifths of a second rather than a chunk at a time. |
Note
🎯 Measured on a 2 minute read: no gap at all. 1.2 s to the first word, and the renderer
stayed ahead of the listener for the remaining 128 s of audio.
🎤 microphone
│
▼
512-sample windows ──► 🧠 Silero VAD ──► speech segments
│
▼
📥 segment queue
│
▼
🔀 decoder thread ──► Parakeet ──► ⌨️ typed
|
The recognizer never sees room noise or keyboard clatter. An early energy-threshold version fired on a single door bump and transcribed nothing. |
Transcribing a phrase takes about as long as saying it. Inline, every word spoken during a decode would be dropped by the buffer. |
Switching dictation off drains the queue first. The last thing you said before hitting the button still lands. |
Important
⚖️ The tradeoff, stated plainly. Vosk streamed words as you spoke them. Parakeet transcribes a whole utterance, so text lands a second or two after you finish a phrase instead of trickling out live.
That is the cost of the accuracy — and the reason the dictation loop was rewritten rather than just repointed at a new model file.
Measured on an 11th-gen Core i7-1185G7 — 4 cores, no discrete GPU:
| Step | Result | |
|---|---|---|
| 🪟 | Window on screen | 2.6 s · waits for nothing but itself |
| 🧩 | Parakeet model load | 3.5 s warm, much longer on a cold disk · behind the open window |
| 🗣️ | Kokoro voice load | ~3 s · after the recognizer, not competing with it |
| 🔊 | First word of a read | 1.2 s · any length of text |
| ⚡ | Decode @ 4 threads | 1.9x realtime ✅ |
| 🐢 | Decode @ 8 threads | 1.1x realtime ❌ |
Note
🧱 What is left is the floor, not a bug. That 3.5 s is 2.4 GB of full precision encoder
weights coming off disk. There are smaller quantized builds; this project does not use one.
It is now spent behind an open window instead of in front of it.
Warning
🧵 More threads is slower. Hyperthread siblings contend for the same execution units. The app derives its own thread count — roughly half the logical processors, clamped to 2–8 — so there is nothing to configure and nothing tuned to one machine.
Download Speak Anywhere for Windows
Unzip it, open the Speak Anywhere folder and run Speak Anywhere.exe. The first time, it
downloads its speech model and voices once (about 3 GB, with a progress bar) into
%LOCALAPPDATA%\SpeakAnywhere\models. After that it starts in seconds and runs offline.
Voice Picker.exe in the same folder lets you audition voices and keep one.
If Windows shows a blue "protected your PC" box, click More info, then Run anyway.
git clone https://github.com/nicedreamzapp/SpeakAnywhere.git
cd SpeakAnywhere
pip install -r requirements.txt⬇️ The models are not in this repo — they're about 3 GB together.
🎧 Download the speech recognizer · 2.5 GB
mkdir -p _resources/parakeet-tdt-0.6b-v2
cd _resources/parakeet-tdt-0.6b-v2
BASE=https://huggingface.co/csukuangfj/sherpa-onnx-nemo-parakeet-tdt-0.6b-v2/resolve/main
curl -LO $BASE/encoder.onnx
curl -LO $BASE/encoder.weights # 2.4 GB, this is the slow one
curl -LO $BASE/decoder.onnx
curl -LO $BASE/joiner.onnx
curl -LO $BASE/tokens.txtencoder.onnx references encoder.weights as external data, so both files must sit in the
same folder.
🎯 There are int8 builds that load faster and use a third of the memory. This project deliberately uses the full-precision export — accuracy is the entire reason for the switch, and quantization is the first thing that erodes it.
🧠 Download the voice activity detector · 630 KB
cd _resources
curl -LO https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx🗣️ Download the text-to-speech voices · 350 MB
cd _resources
curl -LO https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/kokoro-multi-lang-v1_0.tar.bz2
tar -xjf kokoro-multi-lang-v1_0.tar.bz2 && rm kokoro-multi-lang-v1_0.tar.bz2[!CAUTION] 🔢 Take
v1_0, notv1_1. Despite the higher number, v1.1 is the Chinese-focused release — it carries 3 English voices against v1.0's 28.
python speak_anywhere.py| Step | ||
|---|---|---|
| 1️⃣ | Launch | the window is up in about 2 seconds |
| 2️⃣ | 🎤 Tap the mic and talk | tap again to stop |
| 3️⃣ | 🔊 Copy text, tap Speak Clipboard | hear it read back |
| 4️⃣ | 🎚️ Open the voice picker | keep the one you like |
| 5️⃣ | ⏩ Adjust | speed 0.5x–2.0x, mic and speaker both selectable |
| Say | Get |
|---|---|
🗣️ "new line" |
⏎ Enter |
🗣️ "period" |
. |
🗣️ "comma" |
, |
✨ Parakeet emits real punctuation on its own, so these are mostly a holdover for when you want to be explicit.
SpeakAnywhere/
├── 📄 speak_anywhere.py # the application
├── 🗣️ voices.py # voice catalog, saved choice, synthesis
├── 🎚️ voice_picker.pyw # audition voices and pick one
├── 📋 requirements.txt
├── ⚖️ LICENSE.txt
├── 🙏 CREDITS.md
└── 📦 _resources/
├── 🖼️ SpeakAnywhere.ico
├── 🎬 splash_video.mp4
├── 🎵 splash_audio.mp3
├── 🧠 silero_vad.onnx ⬇️ downloaded
├── 🎧 parakeet-tdt-0.6b-v2/ ⬇️ downloaded
├── 🗣️ kokoro-multi-lang-v1_0/ ⬇️ downloaded
└── 🔈 piper/ optional alternate voices
powershell -NoProfile -ExecutionPolicy Bypass -File build-windows.ps1It builds in its own clean environment and writes dist\Speak-Anywhere-windows.zip: one folder
with Speak Anywhere.exe and Voice Picker.exe. The models are left out and fetched on first run.
| Project | Role | |
|---|---|---|
| 🎧 | NVIDIA Parakeet TDT | Speech recognition |
| 🗣️ | Kokoro-82M | Text-to-speech |
| 🧠 | Silero VAD | Voice activity detection |
| ⚙️ | sherpa-onnx | Runtime for all three |
| 🔈 | Piper | Alternate voices |
| 🎚️ | PyAudio · sounddevice | Audio I/O |
| ⌨️ | PyAutoGUI | Keyboard automation |
| 🖥️ | Tkinter | Interface |
| 🎬 | OpenCV · Pygame | Splash video and audio |
Full attribution in CREDITS.md.
Copyright © 2024 Nice Dreamz LLC. All Rights Reserved. Proprietary software — see LICENSE.txt.
Matt Macosko · Nice Dreamz LLC info@nicedreamz.wholesale.com
