GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ArchCoder
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
ArchCoder
Model Tree
openai/whisper-small
Input Modalities
Output Modalities
Supported Functionality
This model adapts Whisper-small to Hindi transcription using low-rank adapters on the decoder's attention projections, leaving the encoder frozen. The goal was to test how much accuracy a small, targeted fine-tune could recover on limited compute and a modest amount of training audio.
openai/whisper-smallq_proj and v_proj in the decoder onlyTraining used 5,122 audio chunks (approximately 10 hours) derived from raw recordings roughly 12 minutes long each. Since Whisper requires inputs under 30 seconds, the raw audio was segmented using timestamped transcripts, with chunks longer than 30 seconds or with fewer than 2 characters of text dropped to avoid out-of-memory errors and noisy padding.
Mel filterbank features were computed one file at a time rather than loading all audio into memory at once, and the processed dataset was cached to the Hugging Face Hub to avoid reprocessing on subsequent sessions.
The companion dataset of extracted features and labels is available at ArchCoder/hindi-whisper-chunks.
q_proj and v_proj, encoder frozengroup_by_length enabled to reduce padding waste, since audio duration varied significantly across the dataset and naive batching would pad every sample to the longest one in the batchTwo implementation issues came up during training. group_by_length requires an explicit length key, which was added from label size, and label smoothing was not compatible with this trainer's decoder-shift path, so labels were kept unmodified instead.
Evaluated against the full Hindi test split of FLEURS (418 samples), with forced_decoder_ids set to lock generation into Hindi transcription mode for both the baseline and fine-tuned model.
| Model | N | WER % | CER % |
|---|---|---|---|
| Whisper-Small (Pretrained) | 418 | 58.93 | 32.04 |
| Whisper-Small (LoRA Fine-tuned) | 418 | 33.41 | 14.70 |
This is a 25.5 point absolute reduction in WER and a 17.3 point absolute reduction in CER over the pretrained baseline, from roughly 10 hours of training data and about 9 minutes of training time on a single free-tier GPU.
python
from transformers import WhisperProcessor, WhisperForConditionalGenerationfrom peft import PeftModelprocessor = WhisperProcessor.from_pretrained("openai/whisper-small", language="hindi", task="transcribe")base_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")model = PeftModel.from_pretrained(base_model, "ArchCoder/whisper-small-hindi-lora")model.generation_config.language = "hindi"model.generation_config.task = "transcribe"
If you use this model, please reference the repository:
markdown
ArchCoder/whisper-small-hindi-lora on Hugging Face