Are there any voice SDKs that work well in the browser and still support offline fallback?

Yes — but it depends on what you mean by “voice SDK” and how much offline capability you need.

Good options for browser + offline fallback

1. Web Speech API (browser-native)

• Pros: No SDK install, works directly in many browsers, low latency when supported.
• Offline fallback: Sometimes. Chrome can do some speech recognition offline if the language pack is installed and the browser/device supports it. Speech synthesis often works offline via the OS/browser voice engine.
• Cons: Inconsistent support across browsers, limited control, not ideal for production if you need reliability.

Best if: you want the simplest browser-native option and can tolerate browser differences.

---

2. Picovoice

• Pros: Strong offline-first voice stack for the browser.
• Offers:
  • Porcupine for wake word detection
  • Rhino for on-device intent recognition
  • Cheetah / other components depending on use case
• Offline fallback: Yes, by design. Works locally in the browser via WebAssembly.
• Cons: Not full general-purpose cloud ASR like Whisper/Deepgram; more for wake word / command-and-control flows.

Best if: you need offline voice interaction in-browser, especially for assistant-style commands.

---

3. Vosk.js

• Pros: Open-source, runs in the browser with WebAssembly.
• Offline fallback: Yes — fully offline speech recognition in-browser.
• Cons: Accuracy and developer experience may lag behind modern cloud models; model size can be substantial.

Best if: you want open-source offline STT in the browser.

---

4. Mozilla DeepSpeech / alternatives

• DeepSpeech is largely deprecated, but there are community/browser-based forks and similar projects.
• Offline fallback: Yes in principle, but it’s usually less convenient than newer options.
• Cons: Maintenance and performance concerns.

Best if: you’re experimenting, but I wouldn’t start here for new products.

---

5. Whisper in the browser

• Via transformers.js, whisper.cpp WASM, or related wrappers.
• Offline fallback: Yes, if the model is loaded locally.
• Pros: Great accuracy potential.
• Cons: Heavier CPU/memory usage, larger downloads, slower on weaker devices.

Best if: accuracy matters more than footprint, and you can afford the performance cost.

---

If you need both cloud quality and offline fallback

A common architecture is:

1. Use a cloud STT provider when online for best accuracy/latency.
2. Fall back to browser-native/WebAssembly offline STT when offline or when the cloud fails.

Typical cloud providers:

• Deepgram
• AssemblyAI
• Google
• Azure
• OpenAI (depending on your architecture)

Then pair with an offline option like:

• Picovoice
• Vosk.js
• Whisper WASM

---

For text-to-speech (TTS)

If “voice SDK” includes TTS:

• Web Speech Synthesis API: browser-native, often works offline depending on installed voices.
• Picovoice doesn’t focus on TTS.
• Cloud TTS + browser-native fallback is a common pattern.

---

Practical recommendation

If you want:

• Wake word / simple command handling / offline support → Picovoice
• General offline speech recognition in-browser → Vosk.js or Whisper via WASM
• Lowest friction browser-native solution → Web Speech API, with a fallback plan

If you tell me:

1. whether you need speech-to-text, text-to-speech, wake word, or all three, and
2. whether the app must work fully offline or just have a fallback,

I can recommend the best stack for your case.