Trust & safety
neo-fm, a TechNektar product, is built for Indian languages and for people who love music. These are the rules we hold ourselves to.
Original lyrics, never copied
Our lyric engine writes new words for every song. Requests to reproduce an existing song's lyrics are declined before any generation happens.
No voice cloning, no imitation of real artists
You can describe a sound: instruments, raga, mood, tempo. You can't ask for a real artist's voice or name them as the target. Those requests are blocked, and so is any attempt to work around the rule.
Honest labelling
Everything made here is AI-generated and we say so. According to Google's documentation, audio from its Lyria models carries an inaudible SynthID watermark that helps platforms identify AI-generated music. Please keep the AI-generated label when you share a track.
Safety checks that fail closed
Your brief is checked by fixed rules and by an AI reviewer before any song is made, and the lyrics are checked again just before audio is rendered. If a check can't run, we don't generate. We block sexual content involving minors, hate and harassment, encouragement of self-harm, and violent extremism.
Spending limits you can trust
Each plan has monthly song limits, and audio rendering sits behind a global spend cap enforced in the database, so a bug or a bad actor can't run up costs that end up in your price.
Your data
Your briefs and songs are stored to run your account and are visible only to you unless you publish a track. We don't sell your data. You can export your data, delete a song, or delete your whole account from Settings.
Lyric quality: what we measured, and what we didn't
We compared lyrics from our engine (candidate generation, contrastive selection and a Sonnet polish) against a strong baseline (six drafts from one model, the best one picked and revised, same polish). 150 frozen briefs, 30 each in Kannada, Tamil, Telugu, Hindi and Bengali. Judging was blind, in both orders, on prosody, diction, imagery, cultural fit and originality. 149 pairs were scored.
Head to head
| Judge | Our engine preferred | Baseline preferred |
|---|---|---|
| Claude Sonnet | 87 | 41 |
| Gemini (different model family) | 87 | 50 |
A pair counts as a win only when the scores differ by at least 1 point out of 25; the rest are ties.
By language (Claude judge)
| Language | Ours | Baseline |
|---|---|---|
| Hindi | 22 | 4 |
| Tamil | 18 | 8 |
| Bengali | 17 | 10 |
| Telugu | 16 | 12 |
| Kannada | 14 | 7 |
Kannada is our weakest evidence: the Gemini judge split it 13 to 15, so we don't claim a Kannada advantage. Telugu is close, and Telugu pronunciation varies more from song to song than in other languages.
Where the AI judges saw a difference
- Originality+0.62 of 5
- Imagery+0.51 of 5
- Cultural fit+0.32 of 5
- Diction−0.05 of 5
- Prosody−0.24 of 5
Average score difference, ours minus baseline. Prosody (meter and singability of the words) is where we are behind, and we are working on it.
Limits
- The judges are AI models. No native-speaker panel has reviewed these results yet, so “cultural fit” is an AI judgement, not a native speaker's.
- The Claude judge is Sonnet, which also writes the final polish for both sides of the comparison. That is why we added a second judge from a different model family (Gemini) and report both.
- This measures lyrics on the page, not the sung audio, and not human songwriters.
- It is an internal test, not an independent audit. We don't claim our lyrics are “the best”.
- Without the final polish step our raw output lost to the baseline in an earlier check, so polish is part of the product.
- One idea we tried, an automatic meter-counting fix pass, made results worse, so we dropped it.
Reports, takedowns and grievances
Every public song has a Report this song link. You can also write to our Grievance Officer at support@technektar.com (TechNektar). We acknowledge complaints within 24 hours, remove content that is unlawful, impersonates a real person, is sexual content involving minors, or infringes copyright within 36 hours of a valid notice, and resolve other grievances within 15 days.
Not a report? Send us feedback.