Model reference · open weights
MERaLiON-SER is an open-weight embedding model from MERaLiON, listed in the AxForge catalogue. AxForge can bring it up on EU-owned hardware for you on request — with the licence handled where one is required.
About
MERaLiON-SER-v1: Multilingual Speech Emotion Model 🎤 Live Demo You can experience speech emotion recognition by selecting Gender/Speech recognition tab via our interactive Hugging Face Space: 👉 MERaLiON-SER Demo Upload an audio clip or record your voice to visualize categorical emotions and dimensional affect trajectories in real time. 📘 Model Summary MERaLiON-SER-v1 is a multilingual speech emotion recognition (SER) model jointly predicting 1. Categorical emotions – 7 discrete classes (Neutral, Happy, Sad, Angry, Surprised, Fearful, Disgusted), and 2. Dimensional affect values – continuous (Valence, Arousal, Dominance). Valence (0 negative, 1 positive), Arousal (0 calm, 1 active), Dominance (0 weak, 1 strong) The design achieves parameter-efficient adaptation for multilingual, paralinguistic affect modeling using just 309 M parameters. Language(s): English (Global & Singapore), Chinese, Malay, Tamil, Limited support for Thai, Indonesian, and Vietnamese. More details on model architecture, training and evaluation is available here: Technical report License: MERaLiON Public License 🎯 Supported Outputs 🧠 Architecture Overview - Modality: Speech-only model - Backbone: Whisper-Medium encoder (frozen) for multilingual acoustic features with LoRA adaptation. The Whisper decoder is frozen and unused. - Downstream: Attention-based pooling + modified ECAPA-TDNN capturing temporal & speaker-invariant cues. - Dual-head outputs: - Categorical (Softmax) → discrete emotion classes. - Dimensional (Sigmoid) → continuous VAD estimation. - Parameter-efficient fine-tuning: LoRA adapters integrated into Q/K/V attention projections. - Objective: Weighted Cross-Entropy + Concordance Correlation Coefficient (CCC) loss. - Pooling: Attention pooling for combining short-term and long-term cues. - Augmentations: MixUp, speed perturbation, and additive noise for robustness. - Model parameters size: 309 M This framework balances computational efficiency, cross-lingual transferability, and robust emotion generalization. 📊 Performance Overview In speech emotion recognition (SER), class imbalance is a persistent challenge, as certain emotions such as neutral or happy typically dominate spontan
Summarised from the published model card. Read the full card on the HuggingFace links below.
Specifications
| Maker | MERaLiON |
|---|---|
| Type | Embedding models |
| Parameters (lead) | 770M |
| Variants | 1 |
| Runs with | transformers |
| Released | 2025-10-21 |
| Popularity | 20k downloads / month |
| Likes | 14 |
| Licence | Commercial licence needed |
How it works
Variants
Open weights ship in several sizes and precisions. One page, all the variants — pick the one that fits your GPU. VRAM figures are estimates from model size.
| Variant | Params | Precision | VRAM | Fits 16 GB | Weights |
|---|---|---|---|---|---|
| MERaLiON-SER-v1 | 770M | BF16 | ~1.8 GB | ✓ | Weights ↗ |
Using it via the API
Once AxForge deploys meralion-ser for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (meralion-ser below is illustrative; you get the exact model name on deployment.)
$ curl -sS https://api.axforge.ai/v1/embeddings \
-H "Authorization: Bearer $AXFORGE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"meralion-ser","input":"text to embed"}'
Licence
The weights are open but its licence needs a commercial agreement for business use. AxForge can arrange that licence and host the model for you — you pay AxForge, we settle with the model’s maker. Ask us for a quote. Read the licence ↗
Explore