<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Haipeng W</title>
    <description>The latest articles on DEV Community by Haipeng W (@haipeng_notes).</description>
    <link>https://dev.to/haipeng_notes</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4255389%2Fb9a72b81-0331-4895-8b37-e17785edc382.png</url>
      <title>DEV Community: Haipeng W</title>
      <link>https://dev.to/haipeng_notes</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/haipeng_notes"/>
    <language>en</language>
    <item>
      <title>I Stopped Fighting Local Whisper Every Time I Needed Talk Notes</title>
      <dc:creator>Haipeng W</dc:creator>
      <pubDate>Fri, 20 Nov 2026 09:43:49 +0000</pubDate>
      <link>https://dev.to/haipeng_notes/i-stopped-fighting-local-whisper-every-time-i-needed-talk-notes-1lhg</link>
      <guid>https://dev.to/haipeng_notes/i-stopped-fighting-local-whisper-every-time-i-needed-talk-notes-1lhg</guid>
      <description>&lt;p&gt;I used to tell myself I would just run Whisper locally after every meetup talk I recorded. Then I would spend twenty minutes remembering which virtualenv still worked, wait half an hour for a forty-minute file, and somehow still end the night with a transcript I had to rebuild into notes. The last time I did that, I closed the laptop, stared at the ceiling for a second, and admitted I wanted a path that started in the browser.&lt;/p&gt;

&lt;h2&gt;
  
  
  The setup that quietly became the project
&lt;/h2&gt;

&lt;h3&gt;
  
  
  A virtualenv is not a note-taking habit
&lt;/h3&gt;

&lt;p&gt;I like local tools, and I still think Whisper is impressive when you want everything on your own machine. That is not the same as having a habit. A habit survives a Thursday when you get home late, you have a blog draft to finish, and you do not want to debug a model download before you can even start listening with your eyes.&lt;/p&gt;

&lt;p&gt;The friction piled up in small, almost reasonable ways. The Python environment had drifted. The first pass did not label speakers, so a panel discussion became a single monologue. I would get accuracy that was already quite good, then spend the time I had saved on glue. At some point the glue became the project, and the write-up did not happen.&lt;/p&gt;

&lt;h3&gt;
  
  
  Accuracy was never the only problem
&lt;/h3&gt;

&lt;p&gt;This is the part I had to unlearn. I kept comparing tools as if the winner would be the one with the cleanest word error rate on a quiet file. In practice, the talks I care about are not quiet. There is a microphone that was "fine," a laugh from the front row, a question that starts before the speaker has finished, and a product name the model has never seen. A slightly better transcript that arrives after I have already lost the evening is not better for my actual job, which is turning a recording into a post, a tutorial, or a set of notes I can share with the people who missed the session.&lt;/p&gt;

&lt;p&gt;I still care about accuracy. I just started caring, maybe more, about whether I would actually finish.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F8nsuo79liceui3tpr6cv.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F8nsuo79liceui3tpr6cv.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I actually do after a talk now
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Drop the file in and leave most of the defaults alone
&lt;/h3&gt;

&lt;p&gt;The workflow I kept is almost too plain to write down, which is probably why I kept it. I upload the recording, or I paste a public talk link when that is what I have, and I let the job run. AudioTranscription works as a browser-based &lt;a href="https://audiotranscription.tools/" rel="noopener noreferrer"&gt;audio to text converter&lt;/a&gt;, and it accepts 30+ audio and video formats, so I stopped extracting tracks first. You can also record a short recap in the browser if all you have is a messy phone file and a need to narrate the useful parts before they fade.&lt;/p&gt;

&lt;p&gt;You can try it without creating an account, which I did on the first night, slightly suspicious. After signing in, there are 240 free minutes a day, which has been enough for meetup recordings and the occasional pairing session I want to mine for a howto. A one-hour file has typically come back in around two minutes, not thirty. I enable speaker labels when there is a Q&amp;amp;A. I add a Vocabulary Group when the talk is full of library names, because those are the words I do not want to correct twenty times. Then I leave it alone and make tea. That last step, I realize, is doing a lot of work.&lt;/p&gt;

&lt;p&gt;On clear recordings, accuracy can reach up to 99%, which is the sort of number I used to treat as marketing until I started measuring my own cleanup time. On a noisy hall mic, I still edit. Technical terms still wander. I am not looking for a tool that makes me uncritical. I am looking for a first draft that does not feel like a dare.&lt;/p&gt;

&lt;h3&gt;
  
  
  Ask the transcript a question instead of scrubbing the timeline
&lt;/h3&gt;

&lt;p&gt;The part that surprised me was not the text itself. It was having the recording and the transcript in one workspace, with timestamps I could click when a line looked optimistic. I used to scrub a waveform hunting for "the bit about connection pooling." Now I search for the phrase, or I ask the built-in chat what the speaker said about it, and then I check the answer against the audio. That last part still matters. The chat is a way in, not a substitute for listening when I am about to quote someone.&lt;/p&gt;

&lt;p&gt;There are also AI summaries, with enough styles that I can ask for notes, a chapter-style recap, or a custom prompt aimed at "steps a reader should try." I treat those as a sketch. Once, a summary smoothed a tentative suggestion into something that sounded like a recommendation. I rewrote it. The mind map is occasionally useful when a talk sprawls. Translation I have used once, to skim a session that was not in the language I write in, and I checked the original before I relied on a sentence.&lt;/p&gt;

&lt;h3&gt;
  
  
  Export, then write in my own words
&lt;/h3&gt;

&lt;p&gt;When the transcript is good enough, I export. Word or TXT if I am drafting a post, SRT or VTT if I am adding captions to a cut of the talk, PDF if I just want something easy to share. The point, for me, is to leave the tool and write. The transcript is raw material. It is not the article. I pull quotes, I fix the names, I throw away the rambling preamble, and I try to make a page that would have helped me if I had been in the back row, half a beat behind.&lt;/p&gt;

&lt;p&gt;This is also where a cleanup pass earns a place, at least sometimes. Filler words are human and, in a transcript, exhausting. Removing them can make a draft more readable without pretending the speaker was a copy editor.&lt;/p&gt;

&lt;h2&gt;
  
  
  The honest split I ended up with
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Local when I want it, the browser when I want to finish
&lt;/h3&gt;

&lt;p&gt;I did not uninstall Whisper. If I have a reason to keep a file entirely local, or I want to tinker, I still can. For the ordinary path, after a weekday talk when I would like to publish notes while I still care, I use the browser. The difference is not philosophical. It is the difference between a workflow I describe and a workflow I repeat.&lt;/p&gt;

&lt;p&gt;If you have also spent an evening fighting an environment just to get a transcript you still have to shape, it might be worth trying the less ceremonial version. Upload the next recording, search it, ask it a question, export it, and write the post in your own voice. The model you run in a terminal can be excellent. My problem was never excellence. It was getting to the writing before the talk had gone cold.&lt;/p&gt;

</description>
      <category>productivity</category>
      <category>ai</category>
      <category>writing</category>
      <category>transcription</category>
    </item>
  </channel>
</rss>
