<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[AI Music Experiment Log]]></title><description><![CDATA[AI Music Experiment Log]]></description><link>https://musicexperimentlog.hashnode.dev</link><image><url>https://cdn.hashnode.com/res/hashnode/image/upload/v1593680282896/kNC7E8IR4.png</url><title>AI Music Experiment Log</title><link>https://musicexperimentlog.hashnode.dev</link></image><generator>RSS for Node</generator><lastBuildDate>Tue, 01 Sep 2026 20:28:19 GMT</lastBuildDate><atom:link href="https://musicexperimentlog.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[Same Japanese Lyrics, Same Settings, 42 Seconds Apart]]></title><description><![CDATA[When an AI music generator returns two songs from one request, it is easy to treat them as interchangeable variations. Their durations alone can show that the system made different structural decision]]></description><link>https://musicexperimentlog.hashnode.dev/same-japanese-lyrics-same-settings-42-seconds-apart</link><guid isPermaLink="true">https://musicexperimentlog.hashnode.dev/same-japanese-lyrics-same-settings-42-seconds-apart</guid><category><![CDATA[Artificial Intelligence]]></category><category><![CDATA[music]]></category><category><![CDATA[Testing]]></category><dc:creator><![CDATA[Elara Vance]]></dc:creator><pubDate>Thu, 27 Aug 2026 14:21:31 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6a9025124ae23b822783bed9/20cf459b-9f6c-4cb6-b2fe-d12ff55fd2e6.jpg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>When an AI music generator returns two songs from one request, it is easy to treat them as interchangeable variations. Their durations alone can show that the system made different structural decisions, even when the lyrics and visible controls did not change.</p>
<p>I ran a small controlled test on August 26, 2026. The goal was narrow: keep one structured set of Japanese lyrics and the visible generation settings fixed, then record the duration of both outputs. This was not a ranking or listening test.</p>
<h2>Test question</h2>
<p>The question was:</p>
<blockquote>
<p>If the lyrics, style description, exclusions, and visible controls are held constant, how different can the durations of two outputs from the same generation task be?</p>
</blockquote>
<p>For this run, I used a browser-based <a href="https://songreelai.com/lyrics-into-song">lyrics into song</a> workflow in lyric-to-music mode. I recorded the inputs before generation and the two returned durations afterward.</p>
<h2>Controlled input</h2>
<p>The Japanese lyric input contained:</p>
<table>
<thead>
<tr>
<th>Input property</th>
<th>Recorded value</th>
</tr>
</thead>
<tbody><tr>
<td>Japanese characters</td>
<td>403</td>
</tr>
<tr>
<td>Lines</td>
<td>41</td>
</tr>
<tr>
<td>Structural labels</td>
<td>8</td>
</tr>
</tbody></table>
<p>The eight labeled sections were:</p>
<pre><code class="language-text">Verse 1
Pre-Chorus
Chorus
Verse 2
Pre-Chorus
Chorus
Bridge
Final Chorus
</code></pre>
<p>The working title was <code>朝焼けの輪郭</code> (<code>Asayake no Rinkaku</code>, or <code>The Outline of Dawn</code>). The labels were included in the lyric text so that the intended high-level structure was explicit rather than left entirely implicit.</p>
<h2>Fixed generation settings</h2>
<p>I kept the following visible settings unchanged for the task:</p>
<table>
<thead>
<tr>
<th>Setting</th>
<th>Value</th>
</tr>
</thead>
<tbody><tr>
<td>Model</td>
<td>V4.5+</td>
</tr>
<tr>
<td>Mode</td>
<td>Lyric to Music</td>
</tr>
<tr>
<td>Instrumental</td>
<td>false</td>
</tr>
<tr>
<td>Vocal gender preference</td>
<td>Female</td>
</tr>
<tr>
<td>Style influence</td>
<td>65%</td>
</tr>
</tbody></table>
<p>The style field was:</p>
<pre><code class="language-text">Japanese city pop, female vocal, warm analog synths,
clean electric guitar, melodic bass, steady 104 BPM,
bittersweet but hopeful, clear Japanese pronunciation,
memorable chorus
</code></pre>
<p>The excluded styles were:</p>
<pre><code class="language-text">heavy metal, trap, aggressive vocals, fast rap
</code></pre>
<p>These controls describe the request, but they do not expose every internal decision made during generation. The test therefore controls the visible inputs, not the generator's hidden state or random sampling.</p>
<h2>Measured result</h2>
<p>The task completed in approximately 174 seconds and used 12 credits. The balance changed from 21,898 to 21,886 credits.</p>
<p>It returned two versions:</p>
<table>
<thead>
<tr>
<th>Output</th>
<th>Duration</th>
<th>Duration in seconds</th>
</tr>
</thead>
<tbody><tr>
<td>Version A</td>
<td>3:43</td>
<td>223</td>
</tr>
<tr>
<td>Version B</td>
<td>4:25</td>
<td>265</td>
</tr>
</tbody></table>
<p>The absolute difference was:</p>
<pre><code class="language-text">265 - 223 = 42 seconds
</code></pre>
<p>Relative to Version A, Version B was approximately 18.8% longer:</p>
<pre><code class="language-text">42 / 223 x 100 = 18.83%
</code></pre>
<p>No lyric edits or visible parameter changes were made between those two outputs because they came from the same task.</p>
<h2>What the result shows</h2>
<p>The strongest conclusion is limited but useful: fixed visible inputs did not determine a fixed output duration in this run.</p>
<p>A 42-second difference is large enough to matter when planning a fixed-length video, estimating edit points, or comparing generations automatically. A workflow that assumes every output from the same lyric sheet will have nearly identical timing would have failed on this pair.</p>
<p>The result also suggests that section labels should be treated as guidance rather than a guaranteed timeline. Even though the input explicitly named two verses, repeated choruses, a bridge, and a final chorus, the two outputs still arrived at different total durations.</p>
<h2>What the result does not show</h2>
<p>This experiment does not establish why Version B was longer. Duration alone cannot tell us whether the difference came from:</p>
<ul>
<li>a longer intro or outro;</li>
<li>repeated or extended sections;</li>
<li>instrumental passages;</li>
<li>different pacing inside a section;</li>
<li>pauses or transitions.</li>
</ul>
<p>I also did not use this run to score pronunciation, vocal quality, mix balance, melodic quality, or adherence to the requested 104 BPM. Those would require listening-based criteria or audio analysis that was outside this test.</p>
<p>Finally, two outputs are not enough to estimate a duration distribution. They show that a meaningful difference occurred, not how frequently similar differences occur.</p>
<h2>A better follow-up protocol</h2>
<p>To extend this into a more informative experiment, I would keep the same lyric sheet and settings, then collect at least ten outputs. For every output I would record:</p>
<ol>
<li>task identifier and generation time;</li>
<li>total duration;</li>
<li>detected section boundaries;</li>
<li>intro and outro length;</li>
<li>whether any labeled section was repeated;</li>
<li>whether all lyric lines appeared;</li>
<li>credits used.</li>
</ol>
<p>With that table, it would be possible to calculate the minimum, maximum, median, and spread of output duration. Listening annotations could then be added as a separate layer instead of being mixed with the objective measurements.</p>
<h2>Practical takeaway</h2>
<p>For a lyrics-to-music workflow, one lyric sheet should be treated as a constraint set, not as a fixed arrangement. If duration matters downstream, generate alternatives first and select an output before committing to video timing, captions, or scene boundaries.</p>
<p>This single run produced songs lasting 223 and 265 seconds from the same recorded input. The 42-second gap is the observation; any explanation for that gap requires another experiment.</p>
]]></content:encoded></item></channel></rss>