coild/whisper_small_kannada_translit_IAST

Whisper Small fine-tuned for Kannada using IAST Romanization via Aksharamukha, addressing token limits in non-Roman scripts.

This is a Whisper Small model fine-tuned for Kannada automatic speech recognition (ASR). The original model has a token limit of 448, which makes it less efficient for non-Roman scripts like Kannada. To improve this, we fine-tuned the model with Romanized Kannada text using IAST (International Alphabet of Sanskrit Transliteration), generated via the Aksharamukha library. This approach has reduced sequence lengths, resulting in a 2x faster inference speed.

Example:

Kannada Tokens: ['à²', '¹', 'à²', '¾', 'à²', '°', 'à³į', 'à²', '¦', 'à²', '¿', 'à²', 'ķ', 'Ġà²', '¶', 'à³', 'ģ', 'à²', 'Ń', 'à²', '¾', 'à²', '¶', 'à²', '¯', 'à²', 'Ĺ', 'à²', '³', 'à³', 'ģ']

Kannada Token Count: 31

IAST Tokens: ['h', 'Äģ', 'rd', 'ika', 'ĠÅĽ', 'ub', 'h', 'Äģ', 'ÅĽ', 'ay', 'ag', 'al', 'Ì', '¤', 'u']

IAST Token Count: 15

Romanized Kannada text uses fewer tokens (15) compared to the original Kannada text (31), resulting in faster processing.

Performance

• Test WER: 28.97%
• Test CER: 5.66%
• Test WER WITH BEAM SEARCH And NORMALIZATION: 23.12%
• Test CER WITH BEAM SEARCH AND NORMALIZATION: 4.95%

Usage

#!pip install whisper_transcriber aksharamukha
from whisper_transcriber import WhisperTranscriber
from aksharamukha import transliterate

# Initialize the transcriber
transcriber = WhisperTranscriber(model_name="coild/whisper_small_kannada_translit_IAST")

# Transcribe an audio file with automatic transcript printing
results = transcriber.transcribe(
    "audio_file.mp3",
    min_segment=25,
    max_segment=30,
    silence_duration=0.2,
    sample_rate=16000,
    batch_size=4,
    normalize=True,
    normalize_text=True,
    verbose=False
)

# Apply transliteration to all results
for segment in results:
    print(f"\n[{segment['start']} --> {segment['end']}]")
    print(transliterate.process('IAST', 'Kannada', segment['transcript']))

Model Details

Model Description

• Developed by: Ranjan Shettigar
• Language(s) (NLP): kn
• Finetuned from model [OpenAI]: whipser-small

• Repository: [More Information Needed]
• Paper [optional]: [More Information Needed]
• Demo [optional]: [More Information Needed]

Training Details

Training and evaluation data

Training Data:

• IISc-MILE Train Set Kannada ASR Corpus
• OpenSLR
• Google/Fleurs Train set

Evaluation Data:

• Google/Fleurs Test Set

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 16
• optimizer: adamw
• epochs: 4

Citation [optional]

BibTeX:

[More Information Needed]