majentik/Gemma-4-E4B-BF16-MERaLiON-Speech-LoRA-SG-MLX

Gemma-4-E4B-BF16 + MERaLiON Speech LoRA for Singapore English (MLX)

A composed Singapore-English ASR model that connects the MERaLiON-3 speech encoder to a BF16 Gemma-4-E4B decoder through a trained projector and rank-16 speech LoRA.

This BF16 release is the recommended quality-first edition: it keeps the decoder in native bfloat16, avoids quantization artifacts, and improves the standalone MERaLiON-3 baseline by 9.69 WER points on the MNSC ASR Part 2 test set.

Acknowledgment and public-release license notice

The development of this product/service was assisted by MERaLiON, an AI model developed by the Agency for Science Technology And Research ("A*STAR").

This model is a derivative work of Gemma 4 (Apache 2.0) and is distributed subject to the Gemma Terms of Use (ai.google.dev/gemma/terms) and the Gemma Prohibited Use Policy (see licenses/gemma-prohibited-use-policy.txt). All downstream users are bound by Gemma Section 3.2 use restrictions, the Gemma Prohibited Use Policy, the MERaLiON-3 Public Licence, and the Singapore Open Data Licence v1 for the MNSC-derived components.

Bundled local license texts include licenses/gemma-prohibited-use-policy.txt, licenses/gemma-4-apache-2.txt, licenses/gemma-terms.txt, licenses/meralion-public-license-v3.txt, and licenses/singapore-open-data-licence-v1.txt.

Important: this is a private full-bundle MLX release, not a vanilla transformers.pipeline checkpoint. Use the bundled runtime/ (or equivalent wiring) to connect speech_encoder/, projector/, decoder/, and lora/. Identity note: This private full-bundle package converts the SG-named adapter repo into a complete MLX bundle using the MNSC-evaluated BF16 full-bundle source. The evaluation rerun is for MNSC ASR Part 2; the SG slug is the repo identifier inherited from the prior public adapter release. No new SG-only evaluation beyond MNSC ASR Part 2 is claimed.

Result summary

Evaluated on MERaLiON Multitask National Speech Corpus v1 — ASR Part 2 Test (3000 utterance-level clips).

• —Absolute improvement vs. MERaLiON-3 baseline: −9.69pp
• —Absolute improvement vs. 8-bit sibling: −2.77pp
• —Normalization: lowercase, ASCII punctuation stripped, whitespace collapsed, speaker-prefix tags removed from reference and hypothesis.

Audience

Currently owner-only/private. A public visibility flip is authorized only if the public-release license gate passes and the user explicitly approves the follow-up visibility change. Do not infer public status from this preparation.

Example outputs

These are actual model outputs from artifacts/run_3000_bf16_r16mlp/eval_predictions.jsonl, selected from the held-out MNSC ASR Part 2 test set. Each row scores 0% WER under the release normalizer (lowercase, punctuation removed, whitespace collapsed).

Across the full saved evaluation file, 1071 / 3000 utterances scored 0% WER, and another 857 scored ≤20% WER under the same normalizer.

What is inside

No MNSC-derived audio sample or sample metadata is bundled. For inference, bring your own 16 kHz mono PCM16 WAV.

The speech path is:

audio -> Whisper-style log-mel -> MERaLiON-3 encoder/adaptor -> 3584-d speech embeddings
      -> projector -> 2560-d Gemma embedding space -> Gemma-4-E4B BF16 + speech LoRA -> text

Quickstart

Download or locate the private full bundle. Bring your own 16 kHz mono PCM16 WAV; no MNSC-derived recording is packaged. Then use the bundled runtime from this repository root:

from pathlib import Path
import sys

from huggingface_hub import snapshot_download

bundle = Path(snapshot_download("majentik/Gemma-4-E4B-BF16-MERaLiON-Speech-LoRA-SG-MLX"))
sys.path.insert(0, str(bundle))

from runtime.inference import load_pipeline, transcribe_with_pipeline

pipeline = load_pipeline(
    meralion_dir=str(bundle / "speech_encoder"),
    gemma_id=str(bundle / "decoder"),
    projector_path=str(bundle / "projector"),
    lora_path=str(bundle / "lora"),
    lora_rank=16,
    lora_target_names=(
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ),
)

text = transcribe_with_pipeline(pipeline, "your_audio.wav", max_tokens=128)
print(text)

Runtime notes:

• —lora_path should point to the directory containing adapters.safetensors (lora/), not to the file itself.
• —The target module list must match the adapter: q/k/v/o/gate/up/down across all 42 decoder layers.
• —Use the prompt Transcribe the following audio: unless you intentionally fine-tune/evaluate a different prompt contract.
• —The speech LoRA is switchable in the runtime: enable speech mode for ASR, disable/scale to 0.0 for plain text generation.

Intended use

Good fits:

• —Singapore English / Singlish automatic speech recognition
• —utterance-level voice notes, routing, search, and agent input
• —MLX-native speech-language research with a shared text decoder

Not intended for:

• —safety-critical or legal/medical transcription
• —diarization, timestamps, speaker identification, or streaming ASR
• —Mandarin-only ASR; a separate switchable Mandarin LoRA is planned

Limitations

• —The LoRA is specialized for Singapore English. Other accents and languages may degrade.
• —Residual errors mostly cluster around rare or ambiguous proper nouns, especially code-switched names and places.
• —Long-form audio was not the optimization target; split long recordings into utterance-sized chunks.
• —This repo is a composed bundle. Generic hub inference widgets will not know how to run it without the elderwise runtime.

Architecture details

• —Speech encoder output dimension: 3584
• —Projector hidden dimension: 3072
• —Decoder embedding dimension: 2560
• —Decoder depth: 42 layers
• —LoRA rank: 16
• —LoRA targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• —Speech-mode LoRA scale used by the release runtime: 20.0

Gemma-4's per-layer embedding side channel is handled in the runtime by supplying explicit per-layer inputs for speech positions instead of forcing speech embeddings through token nearest-neighbor recovery.

License & Attribution

See `PROVENANCE.md` for the full chain of custody. Summary:

• —Decoder: `google/gemma-4-E4B-it`, converted to MLX bfloat16; Gemma Terms, Gemma 4 Apache 2.0 text, and the Gemma Prohibited Use Policy apply.
• —Speech tower: `MERaLiON/MERaLiON-3-10B`; MERaLiON-3 Public Licence applies.
• —Training/evaluation corpus: `MERaLiON/Multitask-National-Speech-Corpus-v1`, derived from IMDA National Speech Corpus / MNSC v1; Singapore Open Data Licence v1 applies to MNSC-derived training/evaluation components and requires attribution without implying official endorsement.
• —Projector + LoRA: trained alignment components for this composition; Gemma Terms as derivative, MERaLiON-3 Public Licence for the speech-derived path, and Singapore Open Data Licence v1 for MNSC-derived training/evaluation components.

Required Gemma notice: Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms.

MERaLiON acknowledgment: The development of this product/service was assisted by MERaLiON, an AI model developed by the Agency for Science Technology And Research ("A*STAR").

This model is a derivative work of Gemma 4 (Apache 2.0) and is distributed subject to the Gemma Terms of Use (ai.google.dev/gemma/terms) and the Gemma Prohibited Use Policy (see licenses/gemma-prohibited-use-policy.txt). All downstream users are bound by Gemma Section 3.2 use restrictions, the Gemma Prohibited Use Policy, the MERaLiON-3 Public Licence, and the Singapore Open Data Licence v1 for the MNSC-derived components.

License archives are bundled in `licenses/`, and the root `NOTICE` records the required notices. No MNSC-derived audio sample is redistributed in this public-release-prep bundle.

The repository remains private until the user explicitly approves a follow-up visibility change after the public-release license gate.

Prominent modified-file notices

The following files were modified from upstream/source staging and are called out for Gemma/Apache/MERaLiON notice compliance:

• —config.json — T6.5 stale repository slug corrected to the full-bundle repo name.
• —speech_encoder/composite_config.json — T6.5 upstream-inherited absolute training paths redacted.
• —speech_encoder/encoder_config.json — T6.5 upstream-inherited absolute training path redacted.
• —README.md — T11 public-release compliance text added, including MERaLiON acknowledgment, Gemma Section 3.2 binding language, license archive links, and no-packaged-audio-sample notice.
• —NOTICE — T11 Gemma, MERaLiON, MNSC, Apache 2.0, and Prohibited Use Policy notices expanded.
• —PROVENANCE.md — T11 modification log and evaluation/sample provenance updated.

Citation

@misc{gemma4_meralion_bf16_speech_lora_mlx_2026,
  title  = {Gemma-4-E4B-BF16 + MERaLiON Speech LoRA for Singapore English (MLX)},
  author = {majentik},
  year   = {2026},
  url    = {https://huggingface.co/majentik/Gemma-4-E4B-BF16-MERaLiON-Speech-LoRA-SG-MLX}
}

Related releases

• —8-bit sibling: `majentik/Gemma-4-E4B-MERaLiON-Speech-LoRA-MNSC-MLX` — smaller, 18.86% WER.
• —This BF16 edition is the recommended release for best transcription quality.