<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Audio on Arash Taher</title>
    <link>https://arashtaher.com/tags/audio/</link>
    <description>Recent content in Audio on Arash Taher</description>
    <generator>Hugo</generator>
    <language>en-us</language>
    <lastBuildDate>Fri, 18 Sep 2026 17:07:00 +0000</lastBuildDate>
    <atom:link href="https://arashtaher.com/tags/audio/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Always-on listening for an offline voice assistant</title>
      <link>https://arashtaher.com/blog/always-on-listening-for-an-offline-voice-assistant/</link>
      <pubDate>Fri, 18 Sep 2026 17:07:00 +0000</pubDate>
      <guid>https://arashtaher.com/blog/always-on-listening-for-an-offline-voice-assistant/</guid>
      <description>&lt;p&gt;This is the second post in the series on creating an offline voice assistant. The &lt;a href=&#34;https://arashtaher.com/blog/learning-local-llms-for-an-offline-voice-assistant/&#34;&gt;first post&lt;/a&gt; covered the learning path. This one covers the initial steps for an always-on listening application.&lt;/p&gt;&#xA;&lt;h2 id=&#34;capture-then-transcribe&#34;&gt;Capture, then transcribe&lt;/h2&gt;&#xA;&lt;p&gt;I started by recording audio in 5-second blocks after a wake word, saving each block as a &lt;code&gt;.wav&lt;/code&gt; file, then sending it to another process to be transcribed with &lt;code&gt;whisper.cpp&lt;/code&gt;. I used &lt;code&gt;PyAudio&lt;/code&gt; for capture and a locally compiled &lt;code&gt;whisper.cpp&lt;/code&gt; binary with the &lt;code&gt;tiny.en&lt;/code&gt; model. The Whisper call was a new process via Python&amp;rsquo;s &lt;code&gt;subprocess&lt;/code&gt; module.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
