MERaLiON SpeechEncoder 2
SpeechEncoder 2 is a 630M-class speech representation model trained from scratch. It extends multilingual coverage beyond the first encoder and provides acoustic features for task-specific training.
Broader multilingual pretraining
Training uses an extended BEST-RQ objective with latent targets and the Muon optimizer. The card includes 60,000 hours of Singapore-accented speech. Supported languages are English, Chinese, Malay, Tamil, Thai, Indonesian and Vietnamese.MERaLiON SpeechEncoder 2 official model card
An encoder, not a completed transcription pipeline
The published configuration identifies MeralionBestRqModel. Its output is a representation of speech. A task head and the relevant training are needed to turn these features into labels or text.MERaLiON SpeechEncoder 2 published configuration
Public weights with encoder-specific terms
SpeechEncoder 2 has its own public-license document. It contains an MIT annex and extra redistribution, rights and liability conditions. The added acknowledgement clause in later AudioLLM licenses should not be assumed to appear in this document.MERaLiON SpeechEncoder 2 Public Licence
Choose the downstream target first
Practical guidance. Plan the task labels and evaluation before training a head. A shared acoustic encoder can support different tasks, but success on one task does not establish success on another. Keep the base encoder and trained head versions together when recording results.MERaLiON SpeechEncoder 2 published configuration
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MERaLiON-SpeechEncoder-2 |
| Training data | 1.4 million hours of unlabelled audio.MERaLiON SpeechEncoder 2 official model card |
| Encoder shape | 24 layers, with 1,024 hidden dimensions.MERaLiON SpeechEncoder 2 published configuration |
| Input representation | 80-dimensional Mel features from 16 kHz audio.MERaLiON SpeechEncoder 2 published configuration |
Related models
- MERaLiON GI v1 / GR
MERaLiON / A*STAR IAIC · MERaLiON Speech Gender Recognition
A two-label speech classifier derived from SpeechEncoder 2, published as GI v1 and described as MERaLiON-GR.
Launch date not confirmedDownloadable - MERaLiON SpeechEncoder 2 ASR-CTC
MERaLiON / A*STAR IAIC · MERaLiON SpeechEncoder ASR
A speech recognizer that fine-tunes SpeechEncoder 2 with supervised audio data and a CTC objective.
Launch date not confirmedAccess not confirmed - 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