Let's Be Honest About Accuracy

Important: Non-English accuracy is lower. The AI models powering transcription (including ours) were primarily trained on English. This means accuracy for other languages is noticeably lower — expect 60-70% for major languages, potentially less for others. You will need to do more editing.

We could tell you our AI works perfectly for every language. It doesn't. Here's the reality:

Better accuracy: Spanish, French, German, Portuguese, Italian — Romance and Germanic languages with Latin script tend to perform reasonably well.

More challenging: Japanese, Korean, Mandarin/Cantonese, Hindi/Tamil, Arabic — different writing systems and tonal languages are harder for the AI.

But here's the thing: for many of these languages, there's nothing available. Lyrics databases are heavily English-focused. Musixmatch, Genius, AZLyrics — they have limited coverage for Asian music especially. So the choice isn't "perfect AI vs. available lyrics." It's "imperfect AI vs. nothing."

Getting 60-70% accurate transcription that you clean up is still dramatically faster than transcribing from scratch by ear — especially for languages you're still learning.

Why Non-English Lyrics Are So Hard to Find

If you've ever searched for lyrics to your favorite J-Pop or K-Pop track, you know the struggle. Western lyrics databases have significant gaps:

• Limited catalog coverage. Musixmatch claims 8+ million songs, but their Asian music coverage is sparse. Many popular J-Pop artists have zero entries. K-Pop is better represented, but still incomplete for older or indie releases.
• Romanization inconsistencies. When lyrics do exist, romanization varies wildly. One site writes "watashi" while another writes "watasi." Korean romanization is even more inconsistent. No standardization means searching is frustrating.
• User-contributed quality. Many non-English lyrics come from fan transcriptions. Quality varies from excellent to completely wrong. Without speaking the language fluently, you can't even tell which is which.
• Synced lyrics are rare. Even when you find lyrics, they're usually static text — no timestamps. If you want LRC files for a Japanese song, you're almost certainly creating them yourself.

The Editing Workflow: Your Secret Weapon

For non-English transcription, the editing tools become essential. Here's how to efficiently clean up AI output:

Find and Replace — This is your most powerful tool for non-English editing. The AI often makes consistent errors — mishearing the same word the same way every time it appears. For example, the AI transcribes "愛してる" (aishiteru) as "愛して" throughout the song. One find/replace fixes all 8 instances in the chorus. What would take 5 minutes manually takes 5 seconds.

Split Lines — The AI doesn't always know where Japanese or Korean sentences naturally break. Split lets you separate lines at the correct points. For example, AI chunks two phrases together because it doesn't recognize the sentence boundary. Click between them → Split → Now they're separate lines with proper timing.

Combine Lines — Sometimes the AI over-splits, creating too many short fragments. Combine merges lines back together. For example, a single Korean phrase is broken into 3 tiny segments. Select them → Combine → One properly-timed line.

Direct Text Editing — Click any line to edit the text directly. If you know the correct lyrics (or can look them up), just type them in. Tip: If you have a partial lyrics sheet, use it as reference while listening. The AI gives you timing; you provide corrections.

Language-Specific Tips

Japanese (日本語)

• AI may mix hiragana/katakana/kanji inconsistently. Decide on your preferred format and use find/replace to standardize.
• Particle errors are common (は vs が, を vs お). These are easy to spot if you know basic Japanese.
• Long vowels sometimes get shortened or doubled incorrectly. Listen carefully for おう/おお patterns.

Korean (한국어)

• Spacing errors are frequent. Korean word boundaries aren't always clear to English-trained AI.
• Similar-sounding syllables get confused (ㅂ vs ㅍ, ㄱ vs ㅋ). Context usually makes the correct choice obvious.
• Idol pronunciation can be stylized — the AI transcribes what it hears, which might not be "proper" Korean.

Spanish (Español)

• Generally good accuracy — Spanish is well-represented in training data.
• Watch for regional variations. Mexican Spanish vs. Spanish Spanish may confuse the AI on certain words.
• Accent marks often get dropped. You may need to add them manually for proper spelling.

Hindi (हिन्दी)

• Devanagari script output varies in quality. Romanized output may be more consistent but loses native readability.
• Bollywood songs often mix Hindi and English — the AI handles code-switching reasonably well.
• Aspirated consonants (क vs ख, ग vs घ) are a common error source.

Setting Realistic Expectations

Good use cases:

• You want a starting point to edit, not perfection
• You know enough of the language to correct errors
• Lyrics don't exist anywhere online
• You need synced/timed lyrics specifically
• You're learning the language and want to study lyrics

Challenging cases:

• You need 100% accuracy with no editing
• You don't know the language at all
• Very heavy accents or dialects
• Extremely fast delivery (some rap)
• Very low-quality source audio

The value proposition: AI gets you 60-80% of the way there. The editing tools help you fix the rest efficiently. Total time is still much less than transcribing from scratch — and you end up with properly synced lyrics that didn't exist before.

FAQ