Voice-to-Text Conversion in Action
Voice-to-text conversion on local machine
Voice-to-Text Conversion in Action
<span className="text-accent">Background</span>
I wanted to verify what exactly the hosts were saying when listening to Japanese radio broadcasts, which led me to think, why not record what they said? This way, I can listen repeatedly, and if possible, convert the audio to text for cross-referencing, which would also help improve my Japanese listening skills. So, I quickly called on Claude code and shared my requirements with him, specifying that it should run on a Mac Apple Silicon system. Initially, I struggled to capture system audio, debugging with him took about an hour, and we got it sorted out.
The UI looks like this:
Although it looks a bit crude, it’s just for personal standalone use, so it doesn't matter. The key goal is to record the audio coming from the system. While there are apps available on the App Store, most of them require payment. Additionally, I'm already spending $200 a month on Claude code subscription, so it would be a waste not to utilize it properly, right? XDD
Once I was able to record, it was time to research how to convert the audio to text. Of course, the most straightforward method is to connect an API directly, sending the audio file and asking it to return the text. This is fine, but if possible, building a local AI model where audio goes in and text comes out would be the best scenario, as it also allows me to learn how to convert voice to text locally.
After asking my friend Blue, I found out that it wasn't too difficult. All I needed to do was download OpenAI’s Whisper model, convert the recording format, and feed it to the local model, which would yield very good results.
The results are as follows:
<Gallery images={[ { src: "https://kazami.me/storage/posts/2025-10-14/zh-tw/1760426199326-3362f3b5-3701-476a-b4d6-35231f25538f.png", alt: "2025-10-14-en-gallery-1" }, { src: "https://kazami.me/storage/posts/2025-10-14/zh-tw/1760426200251-98a50058-000b-447f-b1a7-1578ce4659ea.png", alt: "2025-10-14-en-gallery-2" } ]} />
I wrote a script that automatically converts the recorded audio files and feeds them to the model, instructing it to output them to a specified path. The output can be in three file formats: plain text, src, and vtt.
<Gallery images={[ { src: "https://kazami.me/storage/posts/2025-10-14/zh-tw/1760426394570-be30cafe-1426-49a9-9011-42baf61eaac5.png", alt: "2025-10-14-en-gallery-1" }, { src: "https://kazami.me/storage/posts/2025-10-14/zh-tw/1760426395299-c36e6792-6437-4fbf-b540-4bb18cf3170c.png", alt: "2025-10-14-en-gallery-2" }, { src: "https://kazami.me/storage/posts/2025-10-14/zh-tw/1760426395873-ef1808cd-c7ad-4a86-b5df-c5cc05bc97f9.png", alt: "2025-10-14-en-gallery-3" } ]} />
Previously, a voice post developed by hisohiso used the vtt file format to synchronize the text of the voice message with the audio output.
Actually, this approach was accomplished two months ago. Just a few days ago, I asked myself, how about the reverse? How can text be converted to voice? I did some research, and naturally, the easiest way remains to use an AI API. However, doing it locally is the best fit for me!!! Moreover, I just remembered that my blog supports three languages, so the audio files need to be generated in Traditional Chinese, English, and Japanese as well!
This spontaneous idea sparked the thought that,
<span className="text-[1.35rem] font-semibold leading-[1.65]"><span className="text-sky-500">this blog will start to support voice reading mode from this post</span></span>
! The articles I write will automatically generate audio files in three languages for the viewers to choose their preferences.
Through this research, I've grasped the techniques for converting voice to text and text to voice.