MERaLiON AudioLLM Whisper SEA-LION
The first MERaLiON AudioLLM adapts a speech encoder and a regional text model for audio questions, transcription and other speech tasks. I2R developed it with AI Singapore. It produces text from audio and instructions.
Regional audio training
The card reports 260,000 hours of fine-tuning data across six tasks. Its emphasis is English, including Singapore speech. This is an adapted multimodal model; the upstream speech and text models remain separately credited.MERaLiON AudioLLM Whisper SEA-LION official model card
Downloads and component terms
The older release remains downloadable. Its v1 license includes MIT, SEA-LION and Gemma conditions plus redistribution duties. The license table names Whisper large-v3, but the configuration names large-v2. That document mismatch is left explicit.MERaLiON Public Licence v1, 2024MERaLiON AudioLLM Whisper SEA-LION published configuration
Use as a reproducible earlier baseline
Practical guidance. Use this checkpoint when reproducing early MERaLiON audio experiments. Keep its short-clip preparation and text prompt fixed when comparing later models. Do not assign generation 3 streaming or long-audio service features to these weights.MERaLiON AudioLLM Whisper SEA-LION published configuration
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MERaLiON-AudioLLM-Whisper-SEA-LION |
| Inputs and output | 16 kHz mono audio with text instructions; text output.MERaLiON AudioLLM Whisper SEA-LION official model card |
| Audio guidance | Use clips within 30 seconds.MERaLiON AudioLLM Whisper SEA-LION official model card |
| Speech component | The configuration identifies OpenAI Whisper large-v2.MERaLiON AudioLLM Whisper SEA-LION published configuration |
| Text component | AI Singapore gemma2-9b-cpt-sea-lionv3-instruct, with Gemma 2 ancestry.MERaLiON AudioLLM Whisper SEA-LION published configuration |
Related models
- MERaLiON 2 10B
MERaLiON / A*STAR IAIC · MERaLiON 2
A general audio-understanding model that combines a localized Whisper encoder with Gemma 2 text generation.
Launch date not confirmedDownloadable - MERaLiON SpeechEncoder v1
MERaLiON / A*STAR IAIC · MERaLiON SpeechEncoder
A 630M-class speech encoder trained from scratch with self-supervised learning, focused on English and Singapore speech.
Launch date not confirmedDownloadable