This is a manually curated evaluation protocol for measuring OpenEar quality on Indian English, Hindi, and Hindi-English code switching. Phrases below are meant to be recorded by consenting volunteers during benchmarking; none of them are baked into app behavior.
For each phrase: record in both a quiet room and a moderately noisy room,
using at least two speakers with different English/Hindi accents and speaking
rates. Measure word error rate (WER) against the reference transcripts using
a standard WER script against the whisper.cpp transcription output, and
additionally manually review:
- Code-switching correctness (English words inside Hindi sentences, and vice versa).
- Script choice: Roman Hinglish vs Devanagari. Document which the model picks per sentence; do not promise either.
- Proper nouns ("WhatsApp", "office", names) — high-error categories.
- Punctuation and meaning preservation (no paraphrase).
The same protocol compares tiny, base, small, and — separately — any
quantized variants, noting transcription latency (seconds of audio → finish
time), peak RSS, model load time, and thermal effects after 5 continuous
minutes of dictation.
- "Kal meeting kitne baje start hogi?"
- "Please mujhe woh document WhatsApp par bhej dena."
- "Aaj office thoda late pahunchunga."
- "Is file ko download karke email kar do."
- "Mera phone charge khatam ho gaya hai, charger de dena."
- "Team ne aaj demo clear kar diya, badhiya kaam kiya."
- "Chai chahiye? Sirf do minute mein aa raha hoon."
- "Uska naam spelling kaise hota hai? A-A-R-A-V."
- "Payment pending hai kya? Kal process kara deta hoon."
- "Weekend pe ghar jaa raha hoon, Monday tak wapas."
- "कल मीटिंग कितने बजे है?"
- "मुझे वह document WhatsApp पर भेज देना।"
- "आज office थोड़ा late पहुँचूँगा।"
- "यह file download करके email कर दो।"
- "सुबह दस बजे मिलते हैं, ठीक है?"
- "Kindly do the needful and revert by this evening."
- "My meeting got preponed to eleven o'clock."
- "Where is the charger? My phone has died."
- "The cab is waiting downstairs, please come fast."
- "I will join the call from home tomorrow."
The shipped default model (base) stays the default unless benchmarked
results on two low-end and one mid-range Android device show tiny matches
accuracy within measurement error, in which case tiny becomes default for
low-RAM devices via the "low resource mode" setting.