LiveSTT Adds Parakeet-TDT v3: A New On-Device Speech-to-Text Model for 25 Languages
If you use LiveSTT for offline, on-device speech-to-text transcription, there’s a new model worth knowing about: NVIDIA’s Parakeet-TDT v3 (0.6B). This update adds fast, accurate, multilingual transcription to LiveSTT’s growing lineup of local STT engines — no internet connection, no cloud API calls, and no data leaving your device.
Here’s what’s new, how it compares to the models LiveSTT already supports, and why on-device speech recognition keeps getting better.
What’s New: Parakeet-TDT v3 (0.6B)
Parakeet-TDT v3 is a multilingual speech-to-text model built by NVIDIA. At just 484MB, it packs support for 25 European languages into a compact package that runs entirely on-device:
- Bulgarian
- Croatian
- Czech
- Danish
- Dutch
- English
- Estonian
- Finnish
- French
- German
- Greek
- Hungarian
- Italian
- Latvian
- Lithuanian
- Maltese
- Polish
- Portuguese
- Romanian
- Russian
- Slovak
- Slovenian
- Spanish
- Swedish
- Ukrainian
In hands-on testing, Parakeet-TDT v3 delivers noticeably better English transcription accuracy than LiveSTT’s previous default model, SenseVoiceSmall — making it a strong new option for anyone doing dictation, note-taking, meeting transcription, or voice journaling in a European language.
The Models LiveSTT Already Supports
Parakeet-TDT v3 joins two models already built into LiveSTT, each with a different role.
SenseVoiceSmall
Developed by FunAudioLLM (the Alibaba / FunASR team), SenseVoiceSmall is a lightweight multi-task speech understanding model at 469MB. It supports automatic language identification across Mandarin, Cantonese, English, Japanese, and Korean, making it the go-to choice for East Asian language transcription in LiveSTT.
Apple Speech (Built-In Default)
LiveSTT also ships with Apple’s native Speech framework enabled by default. To be candid: anyone who has used Apple Speech knows its recognition quality can be inconsistent. It’s included for one practical reason — so new users can try LiveSTT’s core features immediately, without first downloading a 400+MB model.
Model Comparison: Which STT Model Should You Use?
| Model | Size | Language Coverage | Best For | On-Device |
|---|---|---|---|---|
| Parakeet-TDT v3 | 484MB | 25 European languages | High-accuracy European language transcription, especially English | Yes |
| SenseVoiceSmall | 469MB | Mandarin, Cantonese, English, Japanese, Korean | East Asian language transcription with auto language ID | Yes |
| Apple Speech | Built-in | Varies by system locale | Instant out-of-the-box trial, no download required | Yes |
Why Apple Speech Still Matters — Even If It’s Not the Best
It’s easy to read Apple’s on-device speech recognition as a weak point, but there’s a more interesting way to look at it: Apple’s gaps in AI capability are exactly what create room for third-party apps like LiveSTT to exist. If Apple’s built-in tools were perfect across the board, independent developers would have far less space to innovate.
That said, credit where it’s due — Apple’s underlying infrastructure is genuinely impressive. CoreML’s optimization is remarkable: a 400+MB speech model runs with only 50–80MB of runtime memory on-device. For mobile apps, where memory pressure directly affects performance and battery life, that efficiency matters enormously. It’s what makes running multiple large, offline STT models on a phone practical in the first place.
What’s Next for LiveSTT
Parakeet-TDT v3’s real-world English accuracy has been strong enough that it’s now edging out SenseVoiceSmall as the preferred model for English transcription within LiveSTT. Looking ahead, the next release will add Parakeet’s Japanese model, further expanding LiveSTT’s multilingual, fully offline transcription capabilities.
Try It Yourself
Whether you’re transcribing meetings in French, dictating notes in German, or just want faster, more accurate offline English speech-to-text, Parakeet-TDT v3 is now available in LiveSTT. Download the model, compare it against SenseVoiceSmall or Apple Speech in your own workflow, and see the difference on-device AI can make.
Download LiveSTT from the App Store.
LiveSTT is an offline, privacy-first speech-to-text app that runs entirely on-device — no cloud processing, no internet required.