Each provider sets its own prices, free tiers and limits, and can change them at any time. The pricing notes here are a guide, last checked . Always confirm the current terms on the provider’s own site before you sign up or switch. We add new providers and models as quickly as we can. Until then, you can enter another model name a provider supports, and we’d love to hear what you’d like next.
- Batch
- Transcribes after you stop recording.
- Real-Time
- Transcribes as you speak and uses that stream for the final text.
- Hybrid
- Shows a live preview, then transcribes the completed recording for the final text. Both passes may be billed.
On your device
Download a model once and dictation works offline, with no per-minute fees. WhisperBubble Edition models are tuned to write cleaner text than the stock versions. Speed depends on the model and your phone.
Parakeet Large
The most accurate English model. Your text appears when you stop. Based on NVIDIA Parakeet Unified 0.6B.
- Languages
- English
- Size
- 663 MB
Nemotron English Real-Time
Your English words appear as you speak. The most accurate live model. Based on NVIDIA Nemotron Speech Streaming English 0.6B.
- Languages
- English
- Size
- 662 MB
Nemotron Multilingual Real-Time
Your words appear as you speak, in the language you choose or one it detects. Based on NVIDIA Nemotron 3.5 ASR Streaming 0.6B.
- Languages
- 32 locales, or automatic detection
- Size
- 682 MB
Parakeet Small
A small, fast English model for older phones or limited storage. Based on NVIDIA Parakeet TDT-CTC 110M.
- Languages
- English
- Size
- 132 MB
Canary Flash
A compact model for Spanish, French and German on older phones. For English, Parakeet Small is smaller and faster. Based on NVIDIA Canary 180M Flash.
- Languages
- English, Spanish, French and German
- Size
- 207 MB
Parakeet Multilingual
English and 24 more European languages, detected automatically. Your text appears when you stop. Based on Moondream’s Parakeet Ultra, retrained from NVIDIA Parakeet TDT 0.6B v3.
- Languages
- English and 24 more European languages
- Size
- 671 MB
SenseVoice Small
A compact, fast model for Chinese, Japanese, Korean and Cantonese, plus English. Based on FunAudioLLM SenseVoice.
- Languages
- English, Chinese, Japanese, Korean and Cantonese
- Size
- 240 MB
Whisper Turbo
Covers the most languages, but it’s slower and the largest download. Based on OpenAI Whisper Turbo.
- Languages
- 99+ languages
- Size
- 1.04 GB
Cloud providers
Add an API key from the provider you choose. Audio goes straight from your phone to them, and they bill your account for what you use. Keys are stored encrypted on your phone, and one key serves every model from that provider.
Pricing notes last checked . Each card links to the provider’s current pricing, which always has the final word.
AssemblyAI
RecommendedAccurate transcription across several models; batch jobs run slower than its live or dictation modes.
New accounts get US$50 in free credit, with no card required. At current prices, that’s over 80 hours of the Dictation API. Current pricing from AssemblyAI
- Dictation APIBatchRecommendedBuilt for dictation, it uploads while you speak and returns cleaned-up text in one call, up to two minutes.
- Universal-3.5 Pro Fast SyncBatchTranscribes short clips, up to two minutes, in one request; longer audio uses async instead.
- Universal-3.5 ProBatchAssemblyAI’s most accurate pre-recorded model, covering 18 languages with automatic fallback beyond that.
- Universal-3.6 Pro Real-TimeReal-TimeAssemblyAI’s recommended real-time model, streaming across 32 languages with built-in diarization.
- Universal-3.5 Pro HybridHybridStreams a Universal-3.6 Pro preview, then finalizes with Universal-3.5 Pro; both passes may be billed.
- Universal-2BatchAssemblyAI’s broadest-language model, used as the fallback when Universal-3.5 Pro doesn’t cover a language.
Groq
RecommendedHosts OpenAI’s open-source Whisper models at low cost, with a free tier that needs no card.
Free, with no card required. Limits apply. Current pricing from Groq
- Whisper Large v3BatchRecommendedGroq’s more accurate Whisper model, at a higher per-hour cost than Whisper Large v3 Turbo.
- Whisper Large v3 TurboBatchFaster and cheaper than Whisper Large v3, with a slightly higher word error rate.
OpenAI
RecommendedOpenAI’s GPT Transcribe models, accurate after you stop or as you speak.
Paid, with a card and prepaid credits. Current pricing from OpenAI
- GPT TranscribeBatchRecommendedOpenAI’s recommended model for completed recordings, and its most accurate here.
- GPT Real-Time TranscribeReal-TimeRecommendedOpenAI’s recommended real-time model; streams text as you speak and may trade accuracy for speed.
- GPT Transcribe HybridHybridShows a live preview, then finalizes from the completed recording; both passes may be billed.
Deepgram
Fast, accurate transcription with the Nova model family, priced per minute of audio.
US$200 in free credits, with no card required. Current pricing from Deepgram
- Nova-3BatchDeepgram’s most accurate general-purpose model, recommended for most transcription use cases.
- Nova-3 Real-TimeReal-TimeStreams Nova-3’s accuracy as you speak; the same stream supplies the final transcript.
- Nova-3 HybridHybridShows a live preview, then finalizes from the completed recording; both passes may be billed.
- Nova-2BatchDeepgram’s lower-cost model, mainly useful now for languages or filler-word detection Nova-3 doesn’t cover.
ElevenLabs
Scribe v2 speech-to-text across 90+ languages, with free and pay-as-you-go tiers.
Free or pay-as-you-go, with an account. Limits apply. Current pricing from ElevenLabs
- Scribe v2BatchElevenLabs’ batch model, built for long and complex recordings across 90+ languages.
- Scribe v2 Real-TimeReal-TimeStreams text with about 150ms latency; optimized for live agents rather than long recordings.
- Scribe v2 HybridHybridShows a live preview, then finalizes from the completed recording; both passes may be billed.
Gladia
Solaria speech-to-text, where Solaria-3 trades language coverage for accuracy on European audio.
€50 in free credits that last a year, then prepaid top-ups. No monthly reset. Current pricing from Gladia
- Solaria-1BatchGladia’s broadest model, covering 100+ languages; Solaria-3 is more accurate on five European languages.
- Solaria-3BatchGladia’s most accurate model on real-world English and European audio, limited to five languages.
- Solaria-1 Real-TimeReal-TimeStreams Solaria-1’s broad language coverage as you speak; the stream supplies the final transcript.
- Solaria-1 + CleanupBatchTranscribes with Solaria-1, then applies your cleanup style in the same job through Gladia.
- Solaria-3 + CleanupBatchThe same cleanup job as Solaria-1 + Cleanup, limited to Solaria-3’s five languages.
Gemini transcription with built-in formatting, speaker labels and disfluency cleanup.
Free tier in many regions. Paid use needs prepaid credits, from US$5. Current pricing from Google
- Gemini 3.5 TranscribeBatchPreviewTranscribes completed recordings with speaker labels, word timestamps and built-in disfluency cleanup.
- Gemini 3.5 Transcribe Real-TimeReal-TimePreviewStreams formatted text as you speak, without the batch model’s speaker labels and timestamps.
- Gemini 3.5 Transcribe HybridHybridPreviewShows a live preview, then finalizes from the completed recording; both passes may be billed.
Meta
Muse Voice Transcribe, Meta’s streaming speech model, at one flat rate for batch or real-time use.
Paid, at US$0.18 per audio hour when last checked, with a card on file. Billed after use. Current pricing from Meta
- Muse Voice Transcribe BatchBatchTranscribes a completed recording at Meta’s flat per-hour rate, the same price as Real-Time.
- Muse Voice Transcribe Real-TimeReal-TimeStreams transcription with built-in speaker labels for 20+ speakers, at the same price as Batch.
Mistral
Voxtral 3 transcription in 50 languages, at a low cost.
Free API access may be available, with billing for higher usage. Current pricing from Mistral
- Voxtral Mini Transcribe 3BatchMistral’s compact model for completed recordings, close to Voxtral Small’s accuracy at a lower cost, across 50 languages.
- Voxtral Small Transcribe 3BatchMistral’s most accurate model for completed recordings, across 50 languages.
- Voxtral Mini Real-Time 3Real-TimeStreams text as you speak in 35+ languages, with configurable latency; the same stream supplies the final transcript, a little less accurate than batch.
- Voxtral Mini 3 HybridHybridShows a live preview, then finalizes from the completed recording; both passes may be billed.
OpenRouter
One key for transcription models from many makers, each billed at that model’s own OpenRouter rate.
Paid, with prepaid OpenRouter credits. The price depends on the model. Current pricing from OpenRouter
- OpenAI GPT TranscribeBatchOpenAI’s current recommended transcription model, routed through OpenRouter.
- NVIDIA Nemotron 3.5 TranscribeBatchNVIDIA’s Nemotron 3.5 transcription model, accurate and economical across 40 languages, routed through OpenRouter.
- AssemblyAI Universal-3.6 ProBatchAssemblyAI’s Universal-3.6 Pro model, routed through OpenRouter.
- Google Gemini 3.5 TranscribeBatchGoogle’s multilingual Gemini 3.5 Transcribe model, routed through OpenRouter.
- ElevenLabs Scribe v2BatchElevenLabs’ multilingual Scribe v2 model, routed through OpenRouter.
- Mistral Voxtral Mini Transcribe 3BatchMistral’s compact Voxtral 3 model, routed through OpenRouter at a lower cost.
- Mistral Voxtral Small Transcribe 3BatchMistral’s larger, more accurate Voxtral 3 model, routed through OpenRouter.
- Meta Muse Voice Transcribe 1.0BatchMeta’s general-purpose Muse Voice Transcribe model, routed through OpenRouter.
- Microsoft MAI-Transcribe 2BatchMicrosoft’s multilingual model, covering 60 languages, routed through OpenRouter.
- Deepgram Nova-3BatchDeepgram’s general-purpose Nova-3 model, routed through OpenRouter.
- OpenAI Whisper Large v3BatchOpen-source Whisper model, more accurate but slower than Whisper Large v3 Turbo.
- OpenAI Whisper Large v3 TurboBatchFaster, cheaper Whisper variant, with a slightly higher error rate than Whisper Large v3.
- Google Chirp 3BatchGoogle’s Chirp 3 model, routed through OpenRouter; not the same model as Gemini 3.5 Transcribe.
- xAI Grok STT 1.0BatchxAI’s first Grok transcription model; xAI has replaced it with Grok Voice Transcribe 2.0 directly.
- NVIDIA Nemotron 3.5 ASR 0.6BBatchNVIDIA’s compact streaming ASR model, covering 40+ languages with configurable latency.
- NVIDIA Parakeet TDT 0.6B v3BatchNVIDIA’s offline Parakeet model for English and 25+ European languages, routed through OpenRouter.
- Qwen3 ASR 1.7BBatchQwen’s larger ASR model, covering 30 languages and 22 Chinese dialects with streaming support.
- Qwen3 ASR 0.6BBatchQwen’s smaller ASR model, with the same 30-language and dialect coverage as the 1.7B version.
- Mistral Voxtral Small 24BBatchAn older, larger Voxtral model from Mistral, routed through OpenRouter.
- Mistral Voxtral Mini 3BBatchAn older, compact Voxtral model from Mistral, routed through OpenRouter.
- Fish Audio Transcribe 1BatchFish Audio’s transcription model, with automatic language detection and optional word alignment.
Soniox
Very low-cost multilingual transcription, consistently among the cheapest providers listed.
Paid usage at a low per-hour price, with an account. Current pricing from Soniox
- STT Async v5BatchSoniox’s file transcription model, across 60+ languages with built-in diarization and formatting.
- STT Real-Time v5Real-TimeSoniox’s streaming model, with the same 60+ language and diarization support as STT Async v5.
- STT v5 HybridHybridShows a live preview, then finalizes from the completed recording; both passes may be billed.
xAI
Grok Voice Transcribe 2.0, fast and accurate; pick a language in settings for formatted numbers and dates.
Paid, with prepaid credits. Batch transcription was US$0.10 per audio hour when last checked. Current pricing from xAI
- Grok Voice Transcribe 2.0BatchxAI’s current transcription model, twice as accurate as version 1.0 at the same price.
- Grok Voice Transcribe 2.0 Real-TimeReal-TimeStreams the same model as Batch, at a higher per-hour rate for live use.
- Grok Voice Transcribe 2.0 HybridHybridShows a live preview, then finalizes from the completed recording; both passes may be billed.
Your own server
Point WhisperBubble at an OpenAI-compatible endpoint on your own computer or server, for transcription, cleanup or both. Enter its base URL or full endpoint URL, and a key only if your server asks for one.
- HTTP or HTTPS. Both work. HTTP doesn’t encrypt audio, text or keys on their way across the network.
- The final address. Connections refuse redirects, so your keys, audio and text never get forwarded somewhere else.
- Any model name. Custom model names transcribe completed recordings (Batch).
Transcript cleanup
Cleanup fixes punctuation, filler words and small mistakes in a finished dictation, in the style you choose, whether light fixes, general polish or email format. It’s off until you turn it on.
In the cloud
Choose from a few approved cleanup models through OpenAI, Google, Groq or OpenRouter, whichever of their providers you prefer. A key you’ve saved for transcription works for cleanup too.
- GPT-6 LunaRecommendedOpenAI’s fast, low-cost model for transcript cleanup; the app’s default pick for new installs. OpenAI OpenRouter
- Gemini 3.5 Flash-LiteGoogle’s fastest, highest-throughput Flash-Lite model, suited to high-volume cleanup. Google OpenRouter
- GPT-OSS 20BOpenAI’s smaller open-weight model on Groq, faster and cheaper than GPT-OSS 120B. Groq
- GPT-OSS 120BOpenAI’s larger open-weight model on Groq, for comparing cleanup quality against GPT-OSS 20B. Groq
AssemblyAI’s Dictation API and Gladia’s + Cleanup models clean up as part of transcription, using your chosen style. You can also add any OpenAI-compatible Chat Completions model, including one on your own server. Cloud cleanup may be billed by the provider.
On your device
For cleanup without a connection, there’s also an experimental model that runs on your phone. The cloud models above give better results.
S1 Mini
By Superwhisper. Cleans up dictation only, and may be slow even on modern phones.
- Languages
- English
- Size
- 484 MB
Missing a model?
Found a great new model or provider that WhisperBubble doesn’t support yet? Send us a note, and we’ll look into adding it.
[email protected]