MSpider3/whisper-hindi2hinglish-swift-ct2

Whisper-Hindi2Hinglish-Swift — CTranslate2 int8

An int8 CTranslate2 conversion of Oriserve/Whisper-Hindi2Hinglish-Swift (a Whisper-base fine-tune by Oriserve for Hindi and Hinglish speech, Apache-2.0), for use with faster-whisper and the ZOLA voice-typing app. All credit for the model goes to Oriserve; this repository only changes the format.

• Output: Hindi and Hinglish speech written in Roman script ("Kal office mein meeting hai, please time par aa jaana"); English stays English.
• Size: 78 MB. Runs on CPU.
• Decode with language="en" — with "hi" the model falls back to Devanagari. zola.json records this for ZOLA.

from faster_whisper import WhisperModel
model = WhisperModel("MSpider3/whisper-hindi2hinglish-swift-ct2", device="cpu", compute_type="int8", cpu_threads=2)
segments, _ = model.transcribe("audio.wav", language="en", beam_size=5, temperature=0.0, vad_filter=True)
print(" ".join(s.text.strip() for s in segments))

Measured on a 2-core laptop CPU (Intel i3-1005G1, 2 threads)

From ZOLA's offline evaluation (synthetic TTS clips plus one real speaker's recording). Word error rate, with Hindi and Hinglish compared after romanization so the output script is not penalized:

	English	Hindi	Hinglish	Real speaker: Hindi / Hinglish	Real-time factor
This model	18%	10%	12%	23% / 13%	0.16
Whisper small (multilingual)	16%	44%	51%	46% / 29%	0.88

These are small test sets; validate on your own audio.

Conversion

ct2-transformers-converter --model Oriserve/Whisper-Hindi2Hinglish-Swift --output_dir whisper-hindi2hinglish-swift-ct2 \
    --quantization int8 --copy_files tokenizer.json preprocessor_config.json

(ctranslate2 4.8.3, transformers 5.20.1)