Echora MOSS 0.9B Multilingual Lyrics Transcriber

A multilingual audio-to-text model for generating lyric text and approximate line timestamps from music audio. No supplied lyric text is required.

Model details

• Model type: MOSS-Transcribe-Diarize fine-tune
• Base model: OpenMOSS-Team/MOSS-Transcribe-Diarize
• Parameters: Approximately 0.9B
• Task: Multilingual lyric transcription with segment timestamps

Intended use

Lyric transcription for personal music libraries and lyric-display projects.

Users are responsible for ensuring they have the necessary rights to process their audio and use the generated text.

Output format

[12.00][S01]Example lyric line[15.20]

Timestamps are expressed in seconds. Speaker labels should not be interpreted as verified singer identities.

Recommended inference

For music recordings, the recommended pipeline uses vocal separation followed by overlapping-window transcription and repetition controls.

These processing steps are separate from the model weights.

Limitations

• Generated lyrics may contain incorrect, omitted, or invented words.
• Repeated phrases, ad-libs, overlapping vocals, and vocal effects may cause repetition or transcription errors.
• Timestamps are approximate.
• Accuracy varies by language, vocal clarity, and recording quality.
• Vocal separation and window merging may introduce additional errors.
• Instrumental sections may produce unintended text.

Review generated lyrics before relying on them. Applications should display a notice such as:

Lyrics transcribed using AI/ML. May contain errors.

Getting started

Model repository:

hcX02/echora-moss-0.9b-multilingual-lyrics-transcriber

Use a compatible MOSS-Transcribe-Diarize inference pipeline to load the model and process audio.

Contact

Open an issue in the Echora repository.