Fine-tuned Hindi ASR model based on
openai/whisper-small, trained
using the Reverse Multi-Stage Fine-Tuning (R-MFT) recipe introduced in the paper
Vividh-ASR: A Complexity-Tiered Benchmark and Optimization Dynamics for Robust Indic Speech Recognition.
This model is part of a set of Malayalam and Hindi Whisper models released by
Adalat AI alongside the Vividh-ASR benchmark.
R-MFT trains in three stages with a decreasing learning rate schedule, presenting
the hardest acoustic data first during the highest-plasticity phase:
Training uses AdamW (weight decay 0.1), linear warmup for the first 10% of
steps, and cosine annealing to zero. Trained on NVIDIA H100 GPUs using
HuggingFace Transformers.
Benchmark WER is measured using
faster-whisper
with 7s VAD segmentation for long-form audio. See the
blogpost for full evaluation details.
1from transformers import pipeline
2
3asr = pipeline(
4 "automatic-speech-recognition",
5 model="adalat-ai/whisper-small-hi-rmft",
6 chunk_length_s=30,
7 device="cuda"
8)
9
10result = asr("audio.wav")
11print(result["text"])
Training data is a superset of the Vividh-ASR benchmark evaluation splits.
Sources used:
This model is intended as a general-purpose Hindi ASR model optimised for
verbatim transcription accuracy across diverse acoustic conditions.
1@misc{vividhasr2025,
2 title = {Vividh-ASR: Diagnosing and Fixing Studio-Bias in Whisper
3 for Indic Languages},
4 author = {[Kush Juvekar, Kavya Manohar, Kumaramanas Nethil]},
5 year = {2026},
6 url = {https://huggingface.co/blog/adalat-ai/vividh-benchmark}
7}
1@misc{vividh2026,
2 title={Vividh-ASR: A Complexity-Tiered Benchmark and Optimization Dynamics for Robust Indic Speech Recognition},
3 author={Kush Juvekar, Kavya Manohar, Aditya Srinivas Menon, Arghya Bhattacharya, Kumarmanas Nethil},
4 year={2026},
5 eprint={2605.13087},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL},
8 url={https://arxiv.org/abs/2605.13087},
9}
