GRT-30M-ASR (MatterWaveAI/grt-30m-asr)

GRT-30M-ASR is a 100% Rotational Speech-to-Text (GRTASRForConditionalGeneration) foundation model (32,594,470 parameters, 0 nn.Linear layers, 0 nn.Conv1d layers) built on the Geometric Rotation Transformer (GRT) architecture and optimized with HessianPairAdam (Sophia-Torus) and GyricAdam.

Raw 16 kHz speech waveforms are transformed via short-time Fourier analysis into log-magnitude/phase polar coordinates r = log(1 + |X|), (r cos φ, r sin φ), rotated through an 11-stage Givens butterfly audio stem (GRTPolarAudioStem), contextualized across 4 bidirectional rotational encoder blocks (GRTAudioModel), folded 2× in time by GRTTemporalRotationalMerger into 64 planar-pair audio tokens, and decoded into text transcriptions by the 12-layer rotational language backbone (GRTModel + GRTCosineHead).

---

Key Highlights & Benchmark Results

Metric	Value
Architecture	GRTASRForConditionalGeneration (model_type = "grt_asr")
Total Parameters	32,594,470 (audio_tower: 3,288,972, language_model: 29,305,498)
nn.Linear / nn.Conv1d Layers	0 / 0 (100% Rotational Purity)
Optimizer	HessianPairAdam (Sophia-Torus) / GyricAdam (#1 in 8-Way Spark Ablation)
Validation Perplexity (Unaligned → Specialized)	29312.46 → 1.0693 (val_loss = 0.0670, 1.0708 w/ GyricAdam)
Transcript Token Top-1 Accuracy	99.59% (GyricAdam / PairAdam)
Phrase Transcription Accuracy	100.0%
Voice Profile Accuracy	100.0%
Prosody Contour Accuracy	95.8% (GyricAdam / PairAdam, 83.3% base)
Full-Sentence Exact Match	95.8% (GyricAdam / PairAdam, 83.3% base)

Held-Out Speech Transcription Samples

Ground-Truth Utterance	GRT-30M-ASR Greedy Transcription
A baritone declarative voice saying hello matter wave.	A baritone calm voice saying hello matter wave.
A baritone declarative voice saying butterfly cascade flow.	A baritone calm voice saying butterfly cascade flow.
A baritone question voice saying quantum phase coherence.	A baritone question voice saying quantum phase coherence.
A baritone excited voice saying hello matter wave.	A baritone excited voice saying hello matter wave.

---

8-Way Classical Geometric Optimizer Ablation (Spark ROCm gfx1151)

We benchmarked all 8 classical geometric, symplectic, and gyroscopic optimizers from transformers.models.grt.optimization_grt on MatterWaveAI/grt-30m-asr (200 steps total per optimizer: 140 Stage-1 rotational audio encoder phase-space alignment steps on GRTAudioModel + 60 Stage-2 causal LM transcript alignment steps on language_model.blocks[-2:] + final_norm, seed = 42, on AMD Radeon 8060S Graphics):

Rank	Optimizer (optimizer_type)	Final Train Loss	Held-Out Val Loss	Held-Out Val PPL	Transcript Token Top-1 Acc	Cross-Audio Margin (ΔL)	Phrase Acc	Prosody Acc	Exact Match	Time (s)
1	GyricAdam (gyric_adam)	0.0925	0.0684	1.0708	99.59%	+1.4590	100.0%	95.8%	95.8%	415.78s
2	PairAdam (pair_adam)	0.1035	0.0875	1.0915	99.59%	+1.4137	100.0%	95.8%	95.8%	215.97s
3	HuygensAdam (huygens_adam)	0.1097	0.0955	1.1002	98.76%	+1.5067	100.0%	87.5%	87.5%	422.14s
4	PolarNaturalAdam (polar_natural_adam)	0.1112	0.0977	1.1026	99.17%	+1.4520	100.0%	91.7%	91.7%	372.97s
5	HessianPairAdam (hessian_pair_adam)	0.1280	0.1072	1.1132	98.76%	+1.4111	100.0%	87.5%	87.5%	219.54s
6	CyclideAdam (cyclide_adam)	0.1827	0.1707	1.1861	96.69%	+1.4527	100.0%	66.7%	66.7%	413.43s
7	HuygensSymplecticAdam (huygens_symplectic_adam)	0.3184	0.2775	1.3198	97.93%	+1.2464	100.0%	79.2%	79.2%	389.00s
8	SymplecticAdam (symplectic_adam)	0.3356	0.3278	1.3879	97.11%	+1.3873	95.8%	70.8%	70.8%	408.80s

Key Geometric Findings on Speech-to-Text (GRTASR)

• Gyroscopic Precession Damping Excels on STFT Phase Alignment (GyricAdam): GyricAdam achieves #1 held-out validation loss (0.0684, val_ppl = 1.0708), #1 transcript token top-1 accuracy (99.59%), and #1 full-sentence greedy exact match (95.8%), outperforming all other optimizers when aligning 1D polar STFT spectrograms through GRTPolarAudioStem and GRTTemporalRotationalMerger.
• Huygens Splay Torque Maximizes Cross-Audio Contrastive Margin (HuygensAdam): HuygensAdam achieves the widest cross-audio transcript discrimination margin (+1.5067 nats) between matched and mismatched audio waveforms while reaching 1.1002 validation PPL (#3 overall).

---

Quickstart (transformers)

import torch
from transformers import AutoTokenizer, GRTASRForConditionalGeneration

model_id = "MatterWaveAI/grt-30m-asr"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = GRTASRForConditionalGeneration.from_pretrained(model_id, torch_dtype=torch.float32).eval()

# 16 kHz speech waveform [1, 20320] (~1.27s)
waveform = torch.randn(1, 20320)
prompt = ("<|video_pad|>" * 64) + "Describe this audio:"
inputs = tokenizer(prompt, return_tensors="pt")

with torch.no_grad():
    out_ids = model.generate(
        **inputs,
        audio_values=waveform,
        max_new_tokens=16,
        do_sample=False,
    )
print(tokenizer.decode(out_ids[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

---

Complete MatterWaveAI/grt-30m* Research Suite on Hugging Face

1. Base Causal LM: MatterWaveAI/grt-30m
2. Chat Fine-Tune: MatterWaveAI/grt-30m-chat
3. 4-Expert Mixture-of-Rotations (MoR): MatterWaveAI/grt-30m-mor
4. Circular-Pooled Sentence Embedding: MatterWaveAI/grt-30m-embedding
5. Cross-Encoder Reranker: MatterWaveAI/grt-30m-reranker
6. Vision-Language (GRTVL): MatterWaveAI/grt-30m-vl
7. Canonical Tokenizer & Polar Embedding Bank: MatterWaveAI/grt-30m-tokenizer
8. Rotational Text-to-Speech (GRTTTS): MatterWaveAI/grt-30m-tts
9. Multimodal Vision-Audio-Language-Speech (GRTOmni): MatterWaveAI/grt-30m-omni
10. Toroidal Rotational LoRA (GRTRotationalLoRA): MatterWaveAI/grt-30m-lora
11. Rotational Speech-to-Text ASR (GRTASR): MatterWaveAI/grt-30m-asr

• Whitepaper: The Geometric Rotation Transformer (GRT): A Purely Rotational Foundation Architecture (GRT_Whitepaper.pdf)

---

License

Apache 2.0