<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
	<channel>
		<title>ASR on itu</title>
		<link>https://itsuitsuki.github.io/tags/asr/</link>
		<description>Recent content in ASR on itu</description>
		<generator>Hugo</generator>
		<language>en</language>
		
		
		
			<copyright>&amp;copy; 2024–2026 itsublog. All rights reserved.
</copyright>
		
		
			<lastBuildDate>Wed, 18 Oct 2023 22:48:39 +0800</lastBuildDate>
		
			<atom:link href="https://itsuitsuki.github.io/tags/asr/index.xml" rel="self" type="application/rss+xml" />
			<item>
				<title>Bengali.AI Speech Recognition: Multi-Module ASR Pipeline</title>
				<link>https://itsuitsuki.github.io/posts/kaggle/bengaliai-speech-2023/</link>
				<pubDate>Wed, 18 Oct 2023 22:48:39 +0800</pubDate>
				<guid>https://itsuitsuki.github.io/posts/kaggle/bengaliai-speech-2023/</guid>
				<description>&lt;p&gt;From July to October 2023, I led a team of 2 in the Kaggle competition &lt;a href=&#34;https://www.kaggle.com/competitions/bengaliai-speech&#34;&gt;Bengali.AI Speech Recognition&lt;/a&gt;, and we won a silver medal, ranking 22nd of 744 teams.&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Our solution writeup on Kaggle, by my teammate Mutian Hong: &lt;a href=&#34;https://www.kaggle.com/competitions/bengaliai-speech/discussion/448119&#34;&gt;22nd Place Solution for the Bengali.AI Speech Recognition Competition&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h2 id=&#34;pipeline&#34;&gt;Pipeline&lt;/h2&gt;&#xA;&lt;p&gt;We developed an automatic speech recognition (ASR) pipeline for a &lt;em&gt;low-resource language&lt;/em&gt; (Bangla/Bengali). With supervised finetuning from augmented low-resource speech datasets, the inference pipeline mainly consists of:&lt;/p&gt;&#xA;&lt;ol&gt;&#xA;&lt;li&gt;denoising;&lt;/li&gt;&#xA;&lt;li&gt;inference with the main ASR decoder model, a finetuned Wav2Vec2, whose decoder probabilities are &lt;em&gt;fused with those of an n-gram draft language model&lt;/em&gt;;&lt;/li&gt;&#xA;&lt;li&gt;post-processing that restores the punctuation of the plain words with a BERT-based &lt;em&gt;punctuation prediction model&lt;/em&gt;.&lt;/li&gt;&#xA;&lt;/ol&gt;&#xA;&lt;h2 id=&#34;my-part&#34;&gt;My part&lt;/h2&gt;&#xA;&lt;p&gt;I worked mainly on the n-gram and punctuation enhancement and on the finetuning of the Wav2Vec2 model.&lt;/p&gt;</description>
			</item>
	</channel>
</rss>
