Whisper-Hindi2Hinglish-Prime-GGML

Whisper Hindi2Hinglish Prime GGML

<!-- builderhelm-header --> <div align="center"> <img src="assets/kanso-hinglish-prime.png" width="160" alt="Kanso Hinglish Prime"/> </div> <div align="center"> <a href="https://getzenvoice.com"><img alt="ZenVoice: Homepage" src="https://img.shields.io/badge/ZenVoice-Homepage-536af5?labelColor=555"/></a> <a href="https://getzenvoice.com/download"><img alt="Download: ZenVoice for Mac" src="https://img.shields.io/badge/Download-ZenVoice%20for%20Mac-536af5?labelColor=555&logo=apple&logoColor=white"/></a> <br> <a href="https://huggingface.co/BuilderHelm"><img alt="Hugging Face: BuilderHelm" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-BuilderHelm-ffc107?labelColor=555"/></a> <br> <a href="#license"><img alt="License: Apache 2.0" src="https://img.shields.io/badge/License-Apache%202.0-f5de53?labelColor=555"/></a> <img alt="Runs on: whisper.cpp" src="https://img.shields.io/badge/Runs%20on-whisper.cpp-8b8b8b?labelColor=555"/> <img alt="Languages: Hindi + English" src="https://img.shields.io/badge/Languages-Hindi%20%2B%20English-8b8b8b?labelColor=555"/> </div> <!-- builderhelm-header -->

Local Hindi and Hinglish speech recognition, with Hindi written in Latin script. Converted by ZenVoice from Oriserve's Whisper-Hindi2Hinglish-Prime, a fine-tune of OpenAI Whisper large-v3. This repository provides a 1.66 GB Q8_0 GGML file for whisper.cpp and ZenVoice.

Model Description

Prime is intended for Hindi and Hindi–English speech with romanized Hindi output. It can preserve English words directly in the transcript, without a separate Devanagari-to-Latin transliteration step. English speech can also produce ordinary English text; output spelling, script and recognition accuracy are not guaranteed.

This is a conversion of Oriserve's existing fine-tune. ZenVoice did not train a new speech model for this repository. The upstream model card describes approximately 550 hours of noisy Indian-accented Hindi training audio.

Architecture

Prime uses the full large-v3 encoder–decoder architecture. Apex is a separate model with a different architecture and tradeoff; this repository does not establish that Prime is universally more accurate than Apex.

Performance

Upstream results, not measurements of this Q8_0 file. Oriserve reports the following WER values using Hindi datasets and Hinglish-format references. Lower is better.

Source: Oriserve's pinned model card. These numbers do not establish WER after GGML conversion and quantization, or accuracy on arbitrary code-mixed dictation.

Recorded conversion smoke checks. The previous version of this repository's card recorded the examples below using Q8_0, whisper.cpp v1.9.1 and Apple Silicon/Metal. They are retained as limited examples, not newly rerun benchmarks.

The final example does not isolate the cause of the error or establish that quantization has no effect. No real-microphone code-mixed WER, throughput, peak-memory or latency benchmark is published for this conversion.

Requirements

Using ZenVoice

Use ZenVoice on a supported Mac and choose the Hinglish language profile with Hinglish Prime. The app's build target is macOS 15 or later; Apple Silicon is the recommended path for this full large-v3 model. Allow room for the 1.66 GB model download plus runtime working memory. File size is not a measure of peak RAM use.

Using whisper.cpp directly

Use a whisper.cpp build that supports Whisper large-v3 and Q8_0 GGML. The recorded checks used v1.9.1; compatibility with every later release is not established here. A C++ toolchain, CMake and Git are needed to build the CLI. The conversion workflow additionally needs Python, PyTorch, Transformers and several GB of disk space for source weights and intermediates. Python is not needed to transcribe with the built CLI and downloaded GGML file.

Installation

ZenVoice

Select Hinglish Prime in ZenVoice's model settings and download it through the app. ZenVoice uses a pinned revision and verifies the expected size and SHA-256 before accepting the model. The engine identifier is hindi2hinglish-prime; the app's display name remains Hinglish Prime.

Direct download

In a Python virtual environment with huggingface_hub installed:

hf download BuilderHelm/Whisper-Hindi2Hinglish-Prime-GGML \
  ggml-hindi2hinglish-prime-q8_0.bin \
  --revision a0c23ffdae030007639da02059256ce4eb1669e8 \
  --local-dir .

echo 'e42f78e509920efc64d99f4ae2b0b39fe59b8f8f37c8bfe41fa6e8607525a2a7  ggml-hindi2hinglish-prime-q8_0.bin' | shasum -a 256 -c -

The old repository name, imYChaudhary22/zenvoice-hinglish-prime-ggml, redirects here. Existing app downloads retain that compatibility alias and the same immutable weights revision. Do not recreate the old repository slug: doing so would interfere with the redirect.

Usage

Build the pinned runtime from source:

git clone https://github.com/ggml-org/whisper.cpp.git
git -C whisper.cpp checkout f049fff95a089aa9969deb009cdd4892b3e74916
cmake -S whisper.cpp -B whisper.cpp/build -DCMAKE_BUILD_TYPE=Release
cmake --build whisper.cpp/build --target whisper-cli whisper-quantize -j

Run from the directory containing the downloaded model. A 16 kHz mono WAV is a straightforward input; convert other recordings with FFmpeg if needed:

ffmpeg -i recording.m4a -ar 16000 -ac 1 -c:a pcm_s16le audio.wav

# Hindi / Hinglish transcription
whisper.cpp/build/bin/whisper-cli \
  -m ggml-hindi2hinglish-prime-q8_0.bin -l hi -f audio.wav

# English transcription: English text is possible
whisper.cpp/build/bin/whisper-cli \
  -m ggml-hindi2hinglish-prime-q8_0.bin -l en -f english.wav

-l supplies a spoken-language hint. It is not a guarantee of output script or a switch that forces romanization. The two hints produced the same output for a few recorded smoke samples; that does not establish that the hint is generally irrelevant. These examples use transcription mode. Adding --translate requests translation to English and changes the task. See the v1.9.1 CLI reference for decoding and output options.

Model Conversion

The conversion preserves the fine-tune's encoder, decoder and vocabulary. It changes storage format and quantizes eligible tensors; it does not add training or a separate romanization model.

Developer workflow, using the whisper.cpp checkout and build above. The OpenAI checkout supplies mel-filter assets required by the upstream converter.

python3.11 -m venv .convert-venv
. .convert-venv/bin/activate
python -m pip install torch 'numpy<2' 'transformers==4.45.2' huggingface_hub

hf download Oriserve/Whisper-Hindi2Hinglish-Prime \
  --revision f85cf467ac948d8114851f7b1cd57a93f8ef2de2 \
  --exclude 'audios/*' \
  --local-dir upstream-prime

git clone https://github.com/openai/whisper.git openai-whisper
git -C openai-whisper checkout 31243bad24cc746f07d4c8bfdd2d974872cb1803
mkdir -p converted-prime
python whisper.cpp/models/convert-h5-to-ggml.py \
  upstream-prime openai-whisper converted-prime

whisper.cpp/build/bin/whisper-quantize \
  converted-prime/ggml-model.bin \
  converted-prime/ggml-hindi2hinglish-prime-q8_0.bin q8_0

This documents the conversion path; it is not a claim that an unpinned Python environment reproduces the published file byte for byte. Compare any rebuilt artifact against the published checksum before treating it as identical. Use the upstream Transformers repository when you need Transformers inference; this repository's .bin file is for whisper.cpp.

File Structure

.
├── README.md                              # This model card
├── ggml-hindi2hinglish-prime-q8_0.bin       # Q8_0 GGML model
└── .gitattributes                         # Hub storage configuration

Model SHA-256:

e42f78e509920efc64d99f4ae2b0b39fe59b8f8f37c8bfe41fa6e8607525a2a7

Limitations

• Romanized Hindi has variable spelling. Exact punctuation, English loanwords and script are not guaranteed.
• Mixed-language speech, accents, noise and proper names can cause omissions or substitutions. The recorded code-mixed TTS failure remains a limitation.
• Whisper can generate unsupported text, especially on silence or poor audio. Review transcripts before using them.
• Upstream WER results do not measure this quantization. Synthetic smoke examples do not substitute for representative microphone recordings.
• This is a full large-v3 model. Download size alone does not predict memory use or speed on your device.
• Translation, diarization and uniform quality across Whisper's other languages are not established by this conversion's checks.

License

Retain applicable license and attribution notices when redistributing the model or runtime. The ZenVoice app has its own repository license.

Citation

Credit Oriserve for the fine-tune, OpenAI for Whisper, and whisper.cpp for the conversion/runtime. For this artifact:

@misc{zenvoice_whisper_hindi2hinglish_prime_ggml,
  title = {Whisper Hindi2Hinglish Prime GGML},
  author = {ZenVoice},
  year = {2026},
  howpublished = {Hugging Face model repository},
  url = {https://huggingface.co/BuilderHelm/Whisper-Hindi2Hinglish-Prime-GGML},
  note = {Q8_0 GGML conversion of Oriserve/Whisper-Hindi2Hinglish-Prime}
}

Acknowledgments

• Oriserve — Hindi/Hinglish fine-tuning and upstream evaluation.
• OpenAI Whisper — base speech model and assets.
• whisper.cpp — native runtime, converter and quantizer.
• ZenVoice — GGML conversion, distribution and app integration.

Published under BuilderHelm AI. Previously imYChaudhary22/Whisper-Hindi2Hinglish-Prime-GGML; existing repository links and pinned revisions redirect to this organization repository. Model filenames and weights are unchanged.