Whisper-Hindi2Hinglish-Prime (whisper.cpp GGML q8_0)
A Hinglish speech-to-text model: speak Hindi, English, or a mix of both — it types romanized Hinglish.
Oriserve/Whisper-Hindi2Hinglish-Prime
converted to whisper.cpp GGML and
quantized to q8_0 for ZenVoice,
an on-device dictation app for macOS.
What it does
| You speak | It types |
|---|---|
| हिंदी (Hindi) | Romanized Hinglish — kya iska svaad achchha hai? |
| Hinglish (Hindi + English mixed) | Romanized Hinglish — mujhe yahi document kal tak complete karna hai |
| English | Romanized Hinglish |
It always writes Latin-script Hinglish. English audio is understood — the vocabulary is English words in romanized spelling — but the output text is Hinglish, never English prose. If you need English text from English audio, use a general Whisper model instead.
Why Prime
Prime is a full Whisper large-v3 fine-tune (1.55B parameters) trained by
Oriserve on ~550 hours of noisy Indian-accented Hindi/Hinglish audio.
Transcribing Hindi speech with a general Whisper model goes through
Devanagari → romanization, which destroys English loanwords:
document → डॉक्यूमेंट → dokyument. Prime writes document directly —
no round trip. It supersedes the turbo-distill
Apex model.
Model file
| Property | Value |
|---|---|
| File | ggml-hindi2hinglish-prime-q8_0.bin |
| Size | 1,656,538,283 bytes (1.66 GB) |
| SHA-256 | e42f78e509920efc64d99f4ae2b0b39fe59b8f8f37c8bfe41fa6e8607525a2a7 |
| Quantization | q8_0 |
| Runtime | whisper.cpp v1.9.1+ |
Provenance
| Stage | Detail |
|---|---|
| Upstream | Oriserve/Whisper-Hindi2Hinglish-Prime @ f85cf467ac948d8114851f7b1cd57a93f8ef2de2 |
| Architecture | Whisper large-v3 — d_model 1280, 32 encoder + 32 decoder layers, 128 mel bins, vocab 51,866 |
| Step 1 — GGML | whisper.cpp/models/convert-h5-to-ggml.py → F16, 3,095,033,483 B |
| Step 2 — quantize | whisper-quantize q8_0 → 1,656,538,283 B |
| Verification | SHA-256 round-trip: downloaded back from this repo and re-hashed, byte-identical |
No weight surgery, no tokenizer changes: the fine-tune keeps the stock 51,866-token Whisper vocabulary, so the standard converter handles it.
Transcription samples
q8_0, whisper.cpp v1.9.1, Apple Silicon (Metal).
| # | Spoken (input audio) | Language hint | Transcribed output |
|---|---|---|---|
| 1 | Hindi — Common Voice common_voice_hi_23796065.mp3 |
hi |
...Kya iska svaad achchha hai? |
| 2 | Hinglish (Indian-accented TTS) — "Mujhe ye document kal tak complete karna hai, aur presentation bhi ready karni hai" | hi |
Mujhe yahi document kal tak complete karna hai aur presentation bhi ready karaani hai. |
| 3 | Same audio | en |
Mujhe yahi document kal tak complete karna hai aur presentation bhi ready karaani hai. |
| 4 | English (TTS) — "Hello, this is a smoke test for the new model integration." | en |
Hello, this is a smoke test for the new model integration. |
| 5 | English (TTS) | hi |
Hello, this is a smoke test for the new model integration. |
| 6 | Code-mixed TTS sentence | en |
Partial — some words garbled |
Rows 4–6 show the model's language hint is mostly irrelevant to the script of its output — it writes Hinglish regardless. Pure Hindi and pure English audio transcribe cleanly; a code-mixed TTS sentence degraded on both F16 and q8_0 (model-inherent, not quantization loss — the upstream card benchmarks pure-Hindi datasets only). Code-mixed accuracy on real Indian-accented speech is unmeasured here; the upstream Speech-To-Text Arena is the reference for comparative quality.
Usage
# Hindi / Hinglish audio (recommended hint)
whisper-cli -m ggml-hindi2hinglish-prime-q8_0.bin -l hi -f audio.wav
# English audio — still outputs Hinglish text
whisper-cli -m ggml-hindi2hinglish-prime-q8_0.bin -l en -f audio.wav
In ZenVoice this model is
offered under the Hinglish language profile and ships as the
hindi2hinglish-prime engine.
Benchmarks (upstream, pure-Hindi datasets)
| Dataset | Whisper Large V3 | Whisper-Hindi2Hinglish-Prime |
|---|---|---|
| Common Voice 2.0 (hi) | 61.94 WER | 32.43 WER |
| FLEURS (hi_in) | 50.84 WER | 28.68 WER |
| Indic-Voices (hi) | 82.56 WER | 60.82 WER |
Numbers from the upstream model card, measured on Hinglish-format references.
Licence
| Component | Licence |
|---|---|
| Upstream model (Oriserve) | Apache-2.0 |
| This conversion repository | Apache-2.0 |
| whisper.cpp runtime | MIT |
Attribution
Whisper-Hindi2Hinglish-Prime by Oriserve, fine-tuned from OpenAI Whisper large-v3. Converted to whisper.cpp GGML and quantized to q8_0 for ZenVoice.
Model tree for imYChaudhary22/zenvoice-hinglish-prime-ggml
Base model
openai/whisper-large-v3