Picovoice Alternative for Small Product Teams
Picovoice's free tier stopped working on June 30, 2026. If your Porcupine wake word just went silent, this page is the migration path — a one-time-payment alternative with no per-device fees.
What happened to Picovoice
Picovoice discontinued its free tier. Free-tier AccessKeys stopped working on June 30, 2026, and the company has confirmed there is no non-commercial tier planned — the product is now enterprise-focused, with a 7-day trial for product teams. Because the SDK validates the AccessKey at engine initialization, even fully local and offline Porcupine deployments (including Home Assistant integrations) stop running without a paid key.
Custom wake-word training through Porcupine had already required a paid account. For a small product team, the question is larger than replacing one SDK: how can the team validate a voice feature, keep control of the model delivery, and avoid building a speech ML group just to ship a short command list?
What a product team is really comparing
| Option | Useful when | Trade-off |
|---|---|---|
| Commercial SDK | You want a packaged vendor runtime | Access keys, platform rules and recurring or deployment fees may apply |
| Voice module | You want hardware and firmware support from one supplier | Requires a hardware change and can increase supplier lock-in |
| Open-source self-training | You have ML and evaluation time | Your team owns datasets, tuning, packaging and device validation |
| Online model platform | You need to validate custom words without building the pipeline | The generated model still needs testing on your target device |
Why Voicute is a possible fit
- Train a custom wake word or fixed command online instead of setting up the training pipeline yourself.
- Test the actual phrase before deciding whether to download the model.
- Receive standard ONNX or TFLite model delivery for compatible targets.
- Use an open offline inference runtime rather than requiring an online activation key at inference time.
- Start with one phrase, then add multiple keywords, real-voice enhancement or false-trigger optimization when the product requires it.
Voicute is not the best choice for every voice product. If the product needs open-ended dictation, names and arbitrary sentences, evaluate ASR instead. If the team needs a fully managed vendor SDK with no model workflow, a commercial SDK may still be the simpler choice.
Why people are migrating
- No continuing personal plan. Picovoice now offers an enterprise trial and directs continuing users to sales.
- Runtime authorization. The SDK requires a valid AccessKey when the engine is initialized.
- Platform-specific artifacts. A custom
.ppnis generated for a selected target platform.
Voicute as the replacement
Voicute is an online custom wake word platform that turns a typed word into a portable, offline ONNX or TFLite model in about 30 minutes. The key difference from Picovoice is the delivered artifact and runtime: the model runs through open inference code without a Voicute AccessKey.
| Picovoice Porcupine | Voicute | |
|---|---|---|
| Free tier | Discontinued (June 30, 2026) | One-time payment per model |
| Continuing personal plan | Not currently offered | One-time model purchase |
| Custom word training | Paid account required | Type a word, auto-train |
| Model artifact | Platform-specific .ppn | Standard ONNX/TFLite (~74–128 KB for current ONNX models) |
| Languages | 9 publicly documented languages | English, German, French, Japanese, Chinese |
| Multi-keyword | Supported | 2–10+ words in one model |
| Runs offline | Yes (until key validation) | Yes, always |
The migration checklist
- Pick your phrase. The same wake word you used with Porcupine, or a new one.
- Train it. Type the word into Voicute; it synthesizes training data and trains the model automatically in about 30 minutes.
- Run it. Download the ONNX file and run it through the open-source View the open-source inference engine and demo models → engine (Apache 2.0) — as a Wyoming service for Home Assistant, or directly on-device.
Python
from wakeword_engine import WakeWordEngine
engine = WakeWordEngine()
engine.load('model_info.json', 'melspectrogram.onnx')
engine.start(lambda word, prob, info: print(f'{word} {prob:.0%}'))FAQ
Is Picovoice's free tier really gone?
Yes. Picovoice confirmed free-tier AccessKeys stopped working on June 30, 2026, with no non-commercial tier planned. Local and offline Porcupine deployments — including Home Assistant — now require a paid key.
What is the best Picovoice alternative for Home Assistant?
For English-only wake words, openWakeWord is the common open-source choice. For custom, non-English, or multi-keyword models without running your own training pipeline, Voicute generates an offline ONNX or TFLite model from a typed word in about 30 minutes and does not require a runtime AccessKey.
How do I migrate from Porcupine to an open alternative?
Generate an ONNX wake-word model for your phrase, then run it through onnx-wakeword as a Wyoming service for Home Assistant, or on-device for Android, Linux, or ESP32.
Does Voicute charge per device?
No. One-time payment per model, and it's yours — no per-device license, no subscription, no per-unit fee.
Can I train a wake word once and use it forever?
Yes. A Voicute model is a self-contained ONNX file (~100–170 KB). After a one-time training payment, deploy it to as many of your own devices as you like, fully offline, with no recurring cost.
Train your custom wake word →