VoxBee was built to run speech recognition on your own device. Whisper and NVIDIA Parakeet models live on your machine, your audio never leaves it, and dictation just works on a plane. That's still the default — and the right answer for most people.
But sometimes you want a hosted model. Maybe you need 100+ languages without downloading a 2.9GB checkpoint. Maybe you need maximum accuracy on a tough recording. Maybe a vendor ships smart formatting that beats the local pipeline. As of v0.6, VoxBee supports 11 cloud speech-to-text providers — all bring-your-own-key, all opt-in. A purple cloud badge stays visible whenever audio is leaving your device, so it never happens by accident.
Here's a practical guide to when each one wins.
The Quick Map
| Provider | Best at | You should try it if |
|---|---|---|
| OpenAI | Familiar quality, smart formatting | You already have an OpenAI key and want a known baseline |
| Deepgram | Latency + cost | You want fast, cheap, English-first dictation |
| AssemblyAI | Long-form accuracy | You're transcribing interviews, podcasts, courses |
| ElevenLabs (Scribe) | Multilingual + diarization | You record across many languages and want speaker labels |
| Groq | Speed | You want Whisper-quality output in a fraction of the wall-clock time |
| xAI Grok | Smart formatting in the X stack | You already live in xAI's ecosystem |
| Mistral Voxtral | European languages | You dictate in French, German, Spanish, Italian, Portuguese |
| Cohere Transcribe | Enterprise-grade English | You're standardized on Cohere for the rest of your stack |
| Speechmatics | Tough accents + noise robustness | You record calls with global English speakers |
| Alibaba Qwen3-ASR | Chinese + Asian languages | You work primarily in Mandarin, Cantonese, Japanese, Korean |
| Soniox | Real-time-quality batch transcription | You want a balance of speed and accuracy with a generous free tier |
How to Set This Up in VoxBee
Open Settings → Cloud Speech. Pick a provider, paste your API key, and choose a model. A "Get API Key" link jumps straight to each provider's dashboard. From v0.6 onward, missing keys, out-of-credits, and model-not-found errors surface as actionable banners with one-click fixes — no more silent failures mid-dictation.
OpenAI — The Familiar Default
OpenAI exposes Whisper-1 and the newer gpt-4o-transcribe family (including a diarized variant). It's the path of least surprise — accuracy is competitive everywhere and smart formatting is on by default. The downside is cost and rate limits compared with specialized providers. Pick OpenAI if you already pay for it and don't want a new dashboard.
Deepgram — Latency and Cost King
Deepgram's nova-3 is purpose-built for English dictation and live transcription. It's noticeably faster than Whisper-class models and meaningfully cheaper per minute. Accuracy is excellent for clean audio. It's not the strongest pick for highly multilingual content, but for English push-to-talk, it's hard to beat.
AssemblyAI — Long-form Accuracy
AssemblyAI's universal-3-pro is tuned for long-form audio with smart formatting and disfluency removal. If you're transcribing hour-long podcasts, interviews, or class recordings, AssemblyAI usually produces a more readable final transcript than raw Whisper. Slower per minute than Deepgram or Groq, but worth it for narrative content.
ElevenLabs Scribe — Multilingual + Diarization
ElevenLabs' scribe_v2 ships built-in speaker diarization and supports a wide range of languages. If you regularly transcribe content with multiple speakers across languages, this is the cleanest option from a single provider.
Groq — Whisper, but Fast
Groq runs OpenAI's Whisper models (large-v3 and large-v3-turbo) on its custom inference hardware. The quality is the same Whisper you know; the wall-clock time is dramatically lower. Great for batch jobs and any time you'd otherwise stare at a spinner.
xAI Grok — Smart Formatting in the X Stack
xAI's grok-stt ships smart formatting by default and slots into the X / Grok ecosystem if that's where your keys already live. Useful if you want to consolidate billing with your existing xAI usage.
Mistral Voxtral — European Languages
Voxtral is Mistral's speech model, and it's especially strong on European languages — French, German, Spanish, Italian, Portuguese. If most of your dictation is in those languages, Voxtral often nudges accuracy above generic English-first models.
Cohere Transcribe — Enterprise English
Cohere is a familiar choice if your organization is already standardized on Cohere for retrieval and chat. Transcribe rounds out the stack with solid English accuracy and consistent billing.
Speechmatics — Tough Accents and Noisy Calls
Speechmatics has been quietly excellent for years on global English accents and noisy phone-quality audio. If you record calls with engineers in Bangalore, customers in Glasgow, and partners in São Paulo, Speechmatics is often the best single pick.
Alibaba Qwen3-ASR — Chinese and Asian Languages
Qwen3-ASR is the right call for Mandarin, Cantonese, Japanese, and Korean work. Western providers have closed much of the gap, but for native-language dictation in CJK languages, Qwen3 still leads.
Soniox — Speed/Accuracy Balance
Soniox is the newcomer in the lineup and surprises people with how good its latency/accuracy curve is. Generous free tier, simple API, very competitive English transcription. Worth a free trial run if you haven't tried it.
What About On-Device?
None of this changes VoxBee's default. Whisper and NVIDIA Parakeet still ship on-device, your audio stays local unless you explicitly choose a cloud provider, and the purple cloud badge keeps you honest. You can switch between on-device and cloud per workflow — fast Parakeet for live dictation, OpenAI for sensitive meeting transcripts, Groq for batch file jobs.
Practical Picks by Workflow
- Daily push-to-talk dictation: Start with Parakeet TDT-CTC 110M on-device. If you want hosted, try Deepgram or Groq.
- Long-form file transcription (podcasts, lectures): Whisper Large v3 on-device, or AssemblyAI / Groq in the cloud.
- Multilingual meetings: ElevenLabs Scribe or OpenAI gpt-4o-transcribe.
- Calls with global English accents: Speechmatics.
- Chinese/Japanese/Korean dictation: Qwen3-ASR.
- European-language dictation: Voxtral.
Getting Started
Download VoxBee (14-day free trial, no account). On-device works out of the box. Open Settings → Cloud Speech to plug in any of the 11 cloud providers with your own API key when you want a hosted model.