May 29, 2026
New features
Model comparison tool (/studio_test)
A new admin workspace for running the same song through multiple model configurations side-by-side, making it easier to evaluate transcription and alignment quality before rolling changes into production.Lyrics validation agent
A second-pass validation agent now reviews transcribed lyrics during the review phase, catching common errors and regenerating transliteration after corrections are applied. Backed by 11 language-specific prompts for higher accuracy across global repertoire.Enhanced history page
The job history view now shows the model used for each run, token usage, estimated cost, and failed-job status — giving you a clearer picture of what each job consumed and why any failures occurred.Updates
Native script enforcement Transcription now enforces native script output for Korean, Japanese, Arabic, and Thai, so thelyrics field stays in the original writing system instead of occasionally returning romanised text.More reliable word-level alignment
Word alignment is more robust on long songs and songs with repeated refrains: higher token limits, automatic JSON repair on truncated responses, and a stronger prompt for handling gaps and refrains. Transient 429 quota errors now retry with backoff instead of failing the job.Bug fixes
- Gemini 500 and 503 errors are now surfaced to users with a clear error state and retry option instead of silently failing.
- Resolved Vercel 504 timeout failures in the transcription and alignment pipeline by raising route limits and adding retry handling.
- Studio now polls for word-level alignment when it isn’t ready at phase entry, eliminating a race that left jobs stuck.
- Fixed an edge case where undefined lyrics lines could break the alignment map.
May 18, 2026
New features
Word-level alignment Jobs now return word-by-word timestamps in addition to line-level timing — enabling true word-highlight karaoke and tighter sync use cases. Word timing is on by default; password_align=false to opt out. See the /transcribe and webhook payloads for the new words_original and words_transliterated fields.New /align endpoint
A dedicated endpoint for distributors who already have lyrics and only need timestamps added. POST /api/v1/align skips transcription entirely, cutting cost and latency. Read the full reference at POST /align.Word-level alignment in transliterated tab
Genuine word-level timestamps are now generated for the transliterated lyrics tab, so romanised playback stays in sync word-for-word with the original.Free trial page at lyrcs.ai/try
Anyone can now run a song through the full karaoke pipeline at lyrcs.ai/try without signing up — a quick way for prospects to evaluate output quality before requesting API access.Updates
Studio rename The word-level alignment experiment has graduated. What was/studio_word is now simply Studio — the production word-level alignment workspace with a sticky audio player across the alignment and review screens.Telugu transcription quality
Added a Telugu corrections dictionary and a diacritic-free romanisation instruction to the transcription prompt, improving accuracy and readability for Telugu lyrics.Cleaner transcription line breaks
Line breaks in transcription output are now more consistent, reducing the amount of manual cleanup needed in the review step.Support email
The contact address has moved from api@lyrcs.ai to support@lyrcs.ai. Existing integrations are unaffected — please update any saved contacts.Bug fixes
- Renamed
/alignresponse fields tolyricsandlyrics_transliteratedfor consistency with the rest of the API. - Fixed the word-level karaoke layout and the word mapping on the transliterated tab.
- Admin pages now render dynamically, resolving intermittent build timeouts.