Downloads · 30 days
165
56% of all-time downloads
svara-whisper-small-hi-ct2
cruqAI/svara-whisper-small-hi-ct2
svara-whisper-small-hi-ct2 is a automatic speech recognition model from cruqAI. Use it when you need speech turned into text. The card lists the license as apache-2.0.
cruqAI/svara-whisper-small-hi converted to CTranslate2 with int8 weights, for CPU inference with faster-whisper or the CTranslate2 Whisper API.
All-time downloads
294
Public
Repo size
248 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.bin248 MB · 98%
At a glance
- Task
- Automatic Speech Recognition
- License
- apache-2.0
- Access
- Public
- Created
- Sep 22, 2026
- Updated
- Sep 22, 2026
- SHA
- 5d98a21d
Base models
Model card
From the Hugging Face model README
svara-whisper-small-hi-ct2
cruqAI/svara-whisper-small-hi converted to CTranslate2 with int8 weights, for CPU inference with faster-whisper or the CTranslate2 Whisper API.
Scored on FLEURS Hindi test (418 utterances) with the same normalisation as the source model's card, on one box:
| WER | CER | |
|---|---|---|
| source model, fp16 on GPU | 0.475 | 0.198 |
| this export, int8 on CPU | 0.489 | 0.205 |
Real-time factor of this export on 4 CPU threads, greedy decoding, feature extraction included: 0.173 (4832 s of audio in 837 s).
from faster_whisper import WhisperModel
model = WhisperModel("cruqAI/svara-whisper-small-hi-ct2", device="cpu", compute_type="int8", cpu_threads=4)
segments, info = model.transcribe("clip.wav", language="hi")
print(" ".join(s.text for s in segments))
Licence and training details are on the source model's card.
- Task
- Automatic Speech Recognition
- License
- apache-2.0
- Languages
- hi
- Created
- Sep 22, 2026
- Updated
- Sep 22, 2026