nagamese-asr / README.md
Badnyal's picture
Update README.md
3bb8fe9 verified
|
Raw History Blame Contribute Delete
3.16 kB
metadata
language:
  - nag
tags:
  - automatic-speech-recognition
  - whisper
  - nagamese
  - low-resource-language
  - speech-recognition
  - asr
license: cc-by-4.0
library_name: transformers
pipeline_tag: automatic-speech-recognition
frameworks:
  - pytorch
metrics:
  - wer
base_model: openai/whisper-small
model_type: whisper
datasets:
  - ARTPARK-IISc
  - Badnyal/nagamese-synthetic-tts
task_categories:
  - speech-recognition
inference: true

ASR Language

Nagamese ASR

Nagamese is a widely used creole language of Northeast India, spoken daily across communities as a bridge of communication.
Despite its reach and importance, it has long remained largely absent from modern speech technology.

This model demonstrates that Nagamese works with contemporary ASR systems-that spoken Nagamese can be recognized and transcribed in practical, real-world settings.

The goal of this project is to create the much needed presence and usability: making Nagamese visible, usable, and supported in today’s AI ecosystem with Nagamese Automatic Speech Recognition (ASR).


Model Details

  • Base model: openai/whisper-small
  • Language: Nagamese
  • Task: Automatic Speech Recognition (ASR)
  • Framework: Transformers (PyTorch)
  • Checkpoint format: safetensors

Training Strategy

Stage 1 – Real Speech Training

  • Dataset: Vaani (real Nagamese speech)
  • ~12k utterances
  • Byte-based audio loading (no HF audio decoding)
  • Result: Strong language understanding, robust to noise

Stage 2 – Synthetic Polishing (Final Model)

  • Dataset: Synthetic Nagamese TTS (~2k samples)
  • Purpose:
    • Improve fluency
    • Fix incomplete word endings
    • Reduce hallucinations and instability
  • Training:
    • Low learning rate
    • Short run (800 steps)
  • Evaluation was always done on real Nagamese speech

This staged approach avoids synthetic bias while improving real-world usability.


Evaluation Notes

  • Raw WER is high due to noise tags and annotation artifacts in real data
  • Qualitative evaluation shows:
    • Correct semantics
    • Improved fluency over Stage‑1
    • Better robustness to spontaneous speech

This model was selected based on human evaluation, not raw WER alone.


Usage

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="MWirelabs/nagamese-asr"
)

asr("audio.wav")

Limitations

  • Designed specifically for Nagamese
  • Not intended for multilingual or code-switched ASR
  • Raw WER should not be used as the sole quality metric

Acknowledgements

  • OpenAI Whisper
  • ARTPARK / IISc Vaani datasets
  • Synthetic speech generated for research purposes

Citation

Citation: https://huggingface.co/MWirelabs/nagamese-asr

@misc{mwirlabs_nagamese_asr_2026, title = {Nagamese ASR}, author = {MWire Labs}, year = {2026}, howpublished = {\url{https://huggingface.co/MWirelabs/nagamese-asr}}, note = {Stage-2 fine-tuned Whisper model for Nagamese speech recognition} }