Offline speech to text on Android, for Indian English and Hinglish

If you speak Indian English, or switch between Hindi and English mid-sentence, you know the problem with most voice typing. Product names come out wrong, Hinglish phrases get “corrected” into nonsense, and every word is sent to a server to be processed.

Phones are now fast enough to do this job themselves. Here is how on-device transcription works and how to get good results from it.

How on-device transcription works

OpenAI’s Whisper is a speech recognition model trained on a very large amount of multilingual audio. The open-source whisper.cpp project runs Whisper efficiently on ordinary hardware, including Android phones. Once a model file is on the phone, audio never needs to leave it.

Picking a model size

Whisper comes in sizes from tiny to large. The trade-off is simple:

• Smaller models are fast and light on battery, and fine for clear speech in a quiet room.
• Larger models are slower but noticeably better with accents, mixed languages, background noise and uncommon words.

A practical approach: start with a mid-size model, and move up if names and mixed-language phrases keep coming out wrong. On a recent phone, the turbo variants of the larger models are a good balance.

Getting names and jargon right

Even a strong model has never heard your colleague’s name or your company’s product. Two techniques fix most of it:

1. Custom vocabulary. Giving the model a list of words to expect nudges it toward them while it decodes.
2. Smart corrections. A small rule list that fixes predictable mistakes after transcription, such as “chat gpt” to ChatGPT or “n eight n” to n8n.

Tips for cleaner transcripts

• Hold the phone close and speak at a normal pace. You do not need to over-enunciate.
• Record in a quiet spot where you can. Voice activity detection helps skip silence, but noise still costs accuracy.
• For long recordings, let punctuation and paragraphing run afterwards rather than dictating “full stop”.

An app built for this

ZwayVoice is our Android app built on exactly these ideas. It records in the background or from the lock screen, transcribes on-device with Whisper, shows text while you speak, and learns your vocabulary and corrections. Everything is searchable, exports to text, Markdown, PDF or JSON, and backs up to your own Google Drive only if you switch that on. There are no analytics, no tracking and no ads.

ZwayVoice is in beta. Request access and tell us which phone you use.