MERaLiON 3 ASR
MERaLiON 3 ASR includes a public 3B-class recognition checkpoint and a separately described hosted service. The model card reports different results for these routes. They are grouped as one ASR release line without claiming that their weights or serving features are identical.
Weights and service versions
The public 3B checkpoint replaces the larger generation 2 ASR baseline in the card's comparisons. The hosted API is described with streaming, timestamps and speaker diarization. Those service capabilities must not be assumed to ship inside the downloadable checkpoint.MERaLiON 3 ASR official model card
The two published duration descriptions differ
The card describes multi-hour decoding for the hosted API. The current OpenAPI transcription route instead states a 30-minute maximum and automatic 30-second chunks. Use the limit documented for the specific route. The sources do not resolve whether a separate route supplies the longer duration.MERaLiON 3 ASR official model cardMERaLiON public OpenAPI specification
Downloads, API keys and speech application
The speech application exposes offline, streaming and streaming-with-diarization modes. The API console requires a key. These interfaces confirm a service offering, not a tested availability or performance guarantee.MERaLiON speech recognition applicationMERaLiON developer API console
The weights use the MERaLiON 3 public license. Hosted requests have separate API terms, including data-processing responsibilities and no guarantee of output accuracy.MERaLiON 3 Public Licence, 2026MERaLiON API terms of use
Evaluate the route you will actually use
Practical guidance. Keep downloadable-model results separate from service results. Use recordings that contain the intended dialects, speaker changes and language switches. Measure transcript quality and speaker-label quality independently; a correct transcript can still have incorrect speaker attribution.MERaLiON 3 ASR published configuration
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable, API, Provider product |
| Model identifiers | MERaLiON-3-3B-ASRMERaLiON-3-ASRMERaLiON-3-ASR-APIMERaLiON-3-3B-ASR-CTM |
| Context window | 8,192 tokens Maximum text-decoder positions in the published configuration; audio consumes representation space and has separate duration guidance. MERaLiON 3 ASR published configuration |
| Download architecture | Whisper large-v3 speech encoder and Gemma 2 2B IT decoder; 26 text-decoder layers.MERaLiON 3 ASR published configuration |
| Regional coverage | English, Mandarin, Malay, Tamil, Indonesian, Thai and Vietnamese, plus Cantonese and Hokkien.MERaLiON 3 ASR official model card |
| Hosted identifier | The current transcription specification names MERaLiON-3-3B-ASR-CTM.MERaLiON public OpenAPI specification |
Related models
- MERaLiON 2 10B ASR
MERaLiON / A*STAR IAIC · MERaLiON 2 ASR
A generation 2 speech recognizer tuned for English and Southeast Asian transcription.
Launch date not confirmedDownloadable - MERaLiON 3 10B
MERaLiON / A*STAR IAIC · MERaLiON 3
A third-generation speech-text model with regional fine-tuning for audio questions, dialogue summaries and contextual analysis.
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