Malayalam ASR

university

Activity Feed

AI & ML interests

None defined yet.

Recent Activity

kavyamanohar authored a paper about 2 months ago

SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR

kavyamanohar authored a paper about 2 months ago

Romanized to Native Malayalam Script Transliteration Using an Encoder-Decoder Framework

kavyamanohar authored a paper about 2 months ago

Vividh-ASR: A Complexity-Tiered Benchmark and Optimization Dynamics for Robust Indic Speech Recognition

View all activity

kavyamanohar

posted an update about 8 hours ago

Post

Some bugs teach you more than they cost you. It's a small bug with a big lesson I keep running into: speech tools are built and benchmarked on English, so their limits are quietly calibrated for English. You only find the edges when you work in the languages they weren't tested on.

Learnt things the hard way, wrote it down so you don't have to.

https://huggingface.co/blog/adalat-ai/whisper-token-limit

kavyamanohar

authored 6 papers about 2 months ago

SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR

Paper • 2605.20712 • Published May 20 • 4

Romanized to Native Malayalam Script Transliteration Using an Encoder-Decoder Framework

Paper • 2412.09957 • Published Dec 13, 2024

Vividh-ASR: A Complexity-Tiered Benchmark and Optimization Dynamics for Robust Indic Speech Recognition

Paper • 2605.13087 • Published May 13 • 3

Multistage Fine-tuning Strategies for Automatic Speech Recognition in Low-resource Languages

Paper • 2411.04573 • Published Nov 7, 2024

What is lost in Normalization? Exploring Pitfalls in Multilingual ASR Model Evaluations

Paper • 2409.02449 • Published Nov 9, 2024

Syllable Subword Tokens for Open Vocabulary Speech Recognition in Malayalam

Paper • 2301.06736 • Published Jan 17, 2023

kavyamanohar

posted an update about 2 months ago

Post

4600

Releasing Vividh-ASR — an open benchmark and models for Hindi and Malayalam ASR.

Vividh-ASR is built from public data, stratified by complexity:
→ Clean recordings
→ Noisy and accented speech
→ Spontaneous, conversational audio

Alongside the benchmark, we release:
→ Open models for Hindi and Malayalam
→ A training recipe with two counterintuitive choices that moved the needle
→ What failed, not just what worked

The stratified evaluation methodology transfers directly to any low-resource language setup — beyond Hindi and Malayalam.

Built at @adalatai , where we build speech tech for Indian courts. This is our first open contribution back to the community. @janaab @Kush0610 @orgh0

Link: https://huggingface.co/blog/adalat-ai/vividh-benchmark