<?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: Build Master</title>
    <description>The latest articles on DEV Community by Build Master (@build_master_5fc54ebe3f3d).</description>
    <link>https://dev.to/build_master_5fc54ebe3f3d</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%2F4254649%2Fc0e0b67a-2598-4e67-a886-a6b2e37d39bb.jpg</url>
      <title>DEV Community: Build Master</title>
      <link>https://dev.to/build_master_5fc54ebe3f3d</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/build_master_5fc54ebe3f3d"/>
    <language>en</language>
    <item>
      <title>What I learned building an audio-to-Suno-v6 prompt tool: language, regional genres and provider outages</title>
      <dc:creator>Build Master</dc:creator>
      <pubDate>Fri, 20 Nov 2026 00:56:20 +0000</pubDate>
      <link>https://dev.to/build_master_5fc54ebe3f3d/what-i-learned-building-an-audio-to-suno-v6-prompt-tool-language-regional-genres-and-provider-1o6k</link>
      <guid>https://dev.to/build_master_5fc54ebe3f3d/what-i-learned-building-an-audio-to-suno-v6-prompt-tool-language-regional-genres-and-provider-1o6k</guid>
      <description>&lt;p&gt;&lt;em&gt;Disclosure: I'm the solo developer of &lt;a href="https://music2prompt.app?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=build-notes" rel="noopener noreferrer"&gt;music2prompt&lt;/a&gt;, the tool this post is about. This post is about what broke, not a sales pitch. Everything in it applies to any audio + LLM project. I drafted this post with help from an AI assistant; the numbers come from my own logs.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Suno v6 launched on September 9, 2026 and retired the older models. Overnight, a lot of people's prompts stopped working the way they used to. The most common request I see is simple: &lt;em&gt;"I have a song I like. Write me the v6 prompt that sounds like it."&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;So I built a tool that listens to a reference track and writes two things: a &lt;strong&gt;Styles&lt;/strong&gt; line and an &lt;strong&gt;Exclude Styles&lt;/strong&gt; list. This is what I learned along the way.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Don't let the model guess the tempo
&lt;/h2&gt;

&lt;p&gt;My first version asked the audio model for everything: genre, mood, instruments, BPM. The genre and instruments were usually good. The BPM often wasn't. It was frequently off by double or half, or rounded to a "typical" tempo for the genre.&lt;/p&gt;

&lt;p&gt;Now tempo is measured in the browser before anything is uploaded. In short:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Decode the audio with the Web Audio API and mix it down to mono.&lt;/li&gt;
&lt;li&gt;Compute a &lt;strong&gt;spectral-flux onset envelope&lt;/strong&gt;: frame-to-frame energy increases across frequency bands.&lt;/li&gt;
&lt;li&gt;Autocorrelate the envelope and score candidate tempos with a small &lt;strong&gt;comb of multiples&lt;/strong&gt; (1×, 2×, 4× the beat period). This avoids the half/double-time trap.&lt;/li&gt;
&lt;li&gt;Refine around the best candidate.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The measured number goes to the model as a hint, and the UI marks it "measured" so the user knows it isn't a guess. The free &lt;a href="https://music2prompt.app/tools/bpm-key-finder/?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=build-notes" rel="noopener noreferrer"&gt;BPM &amp;amp; key finder&lt;/a&gt; uses the same tempo code, plus chroma features and the classic Krumhansl key profiles for the key. One known weakness: relative keys get confused (A minor vs C major), so the key is shown as a suggestion, not a fact.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Lesson:&lt;/strong&gt; if a property can be computed deterministically, compute it, and let the LLM do the parts that need judgment.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. The vocal language matters more than the genre
&lt;/h2&gt;

&lt;p&gt;The biggest quality gap wasn't the genre. It was the language. If you prompt Suno with "Bollywood, tabla, strings" and leave out the language, you often get a song with an English accent. Users notice that immediately.&lt;/p&gt;

&lt;p&gt;So the first field the model must return is the &lt;strong&gt;vocal language&lt;/strong&gt;. "Instrumental" is a valid answer and is treated as one. For vocal tracks the prompt starts with &lt;code&gt;sung in &amp;lt;language&amp;gt;&lt;/code&gt;. Edge cases worth planning for:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Wordless vocals.&lt;/strong&gt; Humming or "ooh-aah" parts aren't a language. Treat them as their own case, or the model may pick one.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Code-switching.&lt;/strong&gt; Spanish-English, Hindi-English and Arabic-French (raï) are common. Both languages belong in the prompt, and the switches belong in the lyric section cues.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dialect.&lt;/strong&gt; "Arabic" isn't enough. Egyptian, Levantine and Gulf Arabic sound very different in Suno, and plain "Arabic" tends to drift toward one of them.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  3. Regional genres need a glossary, not a bigger model
&lt;/h2&gt;

&lt;p&gt;A model can recognize a qawwali or an arabesk song. What it doesn't know is &lt;strong&gt;which words Suno actually responds to&lt;/strong&gt; for that style. Native genre names are often too obscure, and generic words ("Indian music", "Middle Eastern") push Suno toward clichés.&lt;/p&gt;

&lt;p&gt;What worked was a small hand-written &lt;strong&gt;regional genre → Suno tags glossary&lt;/strong&gt; in the system prompt. Each entry has the genre name Suno understands, two or three signature instruments, and a rhythm or tempo anchor. For example:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;qawwali → harmonium, tabla, hand claps, call-and-response chorus, a tempo that builds&lt;/li&gt;
&lt;li&gt;Egyptian mahraganat → maqsoum rhythm, darbuka, heavily autotuned vocals&lt;/li&gt;
&lt;li&gt;Balkan brass → 7/8 meter, trumpets, tuba&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The glossary is small and boring, and it was the most useful part of the system prompt. It also became the seed for a set of &lt;a href="https://music2prompt.app/suno-prompts/?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=build-notes" rel="noopener noreferrer"&gt;genre prompt pages&lt;/a&gt; I keep for 50 regional styles.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Suno v6 changed the output format, so the prompt had to change too
&lt;/h2&gt;

&lt;p&gt;Three v6 rules shaped the output:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Every word in Styles is something to add.&lt;/strong&gt; "no autotune" in Styles can &lt;em&gt;add&lt;/em&gt; autotune. Negatives go in the separate Exclude Styles field, as plain terms.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Long tag lists get averaged.&lt;/strong&gt; About 12–20 dense tags beat 40 vague ones, and the 1,000-character limit is hard.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Section cues inside lyric headers work.&lt;/strong&gt; &lt;code&gt;[Chorus: full band, lead vocal upfront]&lt;/code&gt; does more than &lt;code&gt;[Chorus]&lt;/code&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;I wrote these rules as a small linter that runs in the browser. It's now a free &lt;a href="https://music2prompt.app/prompt-doctor/?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=build-notes" rel="noopener noreferrer"&gt;prompt checker&lt;/a&gt; and a &lt;a href="https://music2prompt.app/suno-v6-prompt-converter/?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=build-notes" rel="noopener noreferrer"&gt;v5.5 → v6 converter&lt;/a&gt;. It's plain JavaScript with regexes and a few lookup tables, and it catches most mistakes before any AI call.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. Your AI provider will go down. Plan for minutes, not seconds
&lt;/h2&gt;

&lt;p&gt;This is the part I wish I'd known earlier. One recent day, about a third of analyses failed. The logs showed:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Every model in the fallback chain failed at the same time&lt;/strong&gt;, with HTTP 500 errors within one second. A fallback chain doesn't help when all the models sit behind the same provider.&lt;/li&gt;
&lt;li&gt;The outages lasted &lt;strong&gt;a few minutes&lt;/strong&gt;. Users who retried by hand 20–60 seconds later usually succeeded.&lt;/li&gt;
&lt;li&gt;One model would sometimes &lt;strong&gt;hang for about 100 seconds&lt;/strong&gt; before timing out. That used up the whole edge-function budget, so the server-side retry never ran.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;What I changed:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;per-model timeout      ~104 s  (p95 of successful calls × 1.5, measured, not guessed)
total server budget     140 s  (edge limit is 150 s — always leave room to refund)
server retry            once, only if every failure was 429/5xx/timeout and &amp;lt; 60 s elapsed
client retry            20 s, then 40 s, with a visible "busy, retrying" notice
credits                 charged once per request, refunded once on final failure
alert                   ≥ 3 failures in 30 min → instant message to my phone
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Two lessons:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Measure before you set timeouts.&lt;/strong&gt; My first guess (40 s) would have cut off about 10% of &lt;em&gt;successful&lt;/em&gt; analyses. The real p95 was 69 s.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep the refund inside the time budget.&lt;/strong&gt; If the platform kills the request at its hard limit, your cleanup code never runs. Then the user loses a credit and you have no log.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The real fix was a second provider path. Analyses now go to Google's Gemini API directly first (a fast model with a 45 s limit, then a second model with a 25 s limit), and the old gateway is only a fallback. Since the switch, analyses take about 10–13 seconds instead of a median of 73. Lesson: a fallback chain only helps if the fallbacks don't share the same point of failure.&lt;/p&gt;

&lt;h2&gt;
  
  
  6. Logging the boring stuff paid off
&lt;/h2&gt;

&lt;p&gt;The most useful change wasn't a feature. It was one text column: &lt;code&gt;error&lt;/code&gt;, holding a compact string like &lt;code&gt;modelA:503 | modelB:429 || retry: modelA:500&lt;/code&gt;. Before it, all I had was "analysis failed". After it, I could see that most failures were whole-provider outages, not bad files. File size and failure rate turned out to be unrelated (r ≈ −0.06).&lt;/p&gt;

&lt;h2&gt;
  
  
  Takeaways
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Compute what you can (tempo, key). Use the LLM for judgment (style, mood, language).&lt;/li&gt;
&lt;li&gt;For non-English music, the &lt;strong&gt;language and dialect&lt;/strong&gt; fields matter more than the genre.&lt;/li&gt;
&lt;li&gt;A small, hand-made &lt;strong&gt;domain glossary&lt;/strong&gt; can beat a bigger model.&lt;/li&gt;
&lt;li&gt;Design for &lt;strong&gt;multi-minute provider outages&lt;/strong&gt;: per-call timeouts, a total budget below the platform limit, client-side retries with backoff, and refunds that always run.&lt;/li&gt;
&lt;li&gt;Log every failed call's status code. You can't fix what you can't see.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you work with Suno and want to try it, the checker and the converter are free and run in your browser. I'm happy to answer questions about the tempo detection or the retry setup in the comments.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>music</category>
      <category>webdev</category>
      <category>llm</category>
    </item>
  </channel>
</rss>
