Choose speed
Parakeet TDT 0.6B v3
Fast and accurate for everyday English dictation, with final text returned quickly on capable phones.
Great when: Quick turnaround matters more than resolving subtle sound-alike words from sentence meaning.
Current Android catalog
Parakeet and Whisper Large-v3 Turbo solve different dictation problems. Loquela lets the phone and the words you speak decide, rather than pretending one model wins every comparison.
Full users can also import custom models.
How to set up offline voice typing on AndroidChoose speed
Fast and accurate for everyday English dictation, with final text returned quickly on capable phones.
Great when: Quick turnaround matters more than resolving subtle sound-alike words from sentence meaning.
Choose context
Sentence-aware multilingual transcription that uses more surrounding meaning, with a longer wait after you stop speaking.
Great when: Sentence meaning needs to resolve names, technical terms, and subtle sound-alike words.
These are capability descriptions, not a universal accuracy ranking. Device performance and audio conditions vary.
Answer three quick questions. Loquela's full catalog shows more detail and checks device compatibility before download.
A starting point
Everyday-phone starting point · Free
Strengths
Tradeoffs
Download: 454 MB
This is a starting suggestion, not a device benchmark. The app checks compatibility before download; memory, speed and accuracy depend on your phone and speech.
Compare models13 models are currently shown for new installs. 42 packages remain verified in the Android catalog so existing legacy installs can still be managed safely.
Versions: 31 (sizes and language packs)
Showing 10 of 13 current choices.
CohereFull
Strength: Auto caps + punctuation
Tradeoff: Very large download (1.7 GB)
Notes: Uses the dictation language you chose; it does not detect it.
Estimated runtime RAM: about 3.2 GB. This is a planning estimate, not a measured device result.
Versions: 2
Hugging FaceFree
Strength: Auto caps + punctuation
Tradeoff: Can repeat/invent words sometimes
Notes: English only.
Estimated runtime RAM: about 549 MB. This is a planning estimate, not a measured device result.
Hugging FaceFull
Strength: Distilled from Whisper Large-v3
Tradeoff: Large download
Notes: English only.
Estimated runtime RAM: about 1.2 GB. This is a planning estimate, not a measured device result.
IBMFull
Strength: Auto caps + punctuation
Tradeoff: Limited real-world track record
Estimated runtime RAM: about 307 MB. This is a planning estimate, not a measured device result.
Versions: 2
Useful SensorsFree
Strength: Very small download
Notes: English only; Batch transcription; Replaces Moonshine v1 Tiny.
Estimated runtime RAM: about 45 MB. This is a planning estimate, not a measured device result.
Useful SensorsFree
Strength: Higher-capacity Moonshine v2
Notes: English only; Batch transcription; Replaces Moonshine v1 Base.
Estimated runtime RAM: about 166 MB. This is a planning estimate, not a measured device result.
NVIDIAFull
Strength: Live on-device text
Tradeoff: Large download
Notes: 560 ms chunk; 19 transcription-ready and 13 broad-coverage locales; Eight additional tokenizer locales require fine-tuning and are not offered; Greedy decoding only.
Estimated runtime RAM: about 712 MB. This is a planning estimate, not a measured device result.
Versions: 2
MetaFree
Strength: Covers 1,600+ languages
Tradeoff: No punctuation or capitals
Estimated runtime RAM: about 461 MB. This is a planning estimate, not a measured device result.
MetaFull
Strength: Covers 1,600+ languages
Tradeoff: No punctuation or capitals
Estimated runtime RAM: about 1.3 GB. This is a planning estimate, not a measured device result.
Versions: 4
NVIDIAFree
Strength: Fast and accurate
Tradeoff: Can confuse similar-sounding names or terms
Notes: English only.
Estimated runtime RAM: about 162 MB. This is a planning estimate, not a measured device result.
NVIDIAFree
Strength: Fast multilingual transcription
Tradeoff: Large download
Estimated runtime RAM: about 730 MB. This is a planning estimate, not a measured device result.
NVIDIAFree
Strength: Fast and accurate
Tradeoff: Large download
Notes: English only.
Estimated runtime RAM: about 723 MB. This is a planning estimate, not a measured device result.
NVIDIAFree
Strength: Newest English Parakeet
Tradeoff: Large download
Notes: English only.
Estimated runtime RAM: about 796 MB. This is a planning estimate, not a measured device result.
QwenFree
Strength: Strong early accuracy benchmarks
Tradeoff: Large download
Notes: Very new, limited track record.
Estimated runtime RAM: about 1.3 GB. This is a planning estimate, not a measured device result.
Alpha CepheiFree
Strength: Very small download
Tradeoff: Lower accuracy
Estimated runtime RAM: about 300 MB. This is a planning estimate, not a measured device result.
OpenAIFull
Strength:
Tradeoff: Repeats words a bit more than full Large
Estimated runtime RAM: about 846 MB. This is a planning estimate, not a measured device result.