GRT-30M-TTS (MatterWaveAI/grt-30m-tts)

GRT-30M-TTS is a 100% Native Rotational Text-to-Speech & Speech-Equipped Language Model (29,352,859 total parameters, zero nn.Linear layers, zero nn.Conv1d layers) built on top of MatterWaveAI/grt-30m.

Architecture Summary

Component	Module	Parameters	Geometry
Language Backbone	GRTModel + GRTCosineHead	29,305,498	12-layer causal Givens Rotational Transformer (MatterWaveAI/grt-30m, frozen to preserve 100% text fluency)
Polar Vocoder Head	GRTPolarVocoderHead (model.speech_head)	47,361	15-stage Givens butterfly expand → GRTRadialGELU → 15-stage butterfly contract (frames_per_token=4, 128 (Re, Im) STFT bins → torch.istft @ 16 kHz)
Total (GRT-30M-TTS)	GRTTTSForConditionalGeneration	29,352,859	0 nn.Linear, 0 nn.Conv1d

Speech Synthesis Alignment Metrics (HessianPairAdam / Sophia-Torus)

• Frame Polar Cosine Similarity: -0.0779 (unaligned) → 0.9874 (aligned)
• Utterance Polar Cosine Similarity: -0.0772 → 0.9684
• Magnitude Envelope Cosine Similarity: 0.9843
• Spectral SNR: -6.99 dB → 12.03 dB

---

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-tts (GRTPolarVocoderHead + 4-codebook RVQ polar cosine codec bank GRTRVQCodecBank, K = 64 per codebook, 200 steps per optimizer, batch_size = 64, seed = 42, on AMD Radeon 8060S Graphics):

Rank	Optimizer (optimizer_type)	Final Train Loss	Held-Out RVQ Val Loss	Held-Out RVQ Val PPL	RVQ Mean Top-1 Acc (4 Codebooks)	Codebook-0 Primary Acc	Text-Conditioned Margin (ΔL)	Frame Polar Cosine Sim	Spectral SNR (dB)	Time (s)
1	HessianPairAdam (hessian_pair_adam)	1.6086	1.6574	5.2456	42.31%	40.30%	+0.1942	0.9786	11.27 dB	28.47s
2	PolarNaturalAdam (polar_natural_adam)	1.6124	1.6606	5.2626	42.37%	40.24%	+0.1912	0.9781	11.24 dB	25.97s
3	PairAdam (pair_adam)	1.6146	1.6647	5.2841	42.14%	39.86%	+0.1785	0.9788	11.37 dB	26.17s
4	HuygensAdam (huygens_adam)	1.6178	1.6661	5.2917	42.36%	40.02%	+0.1834	0.9783	11.22 dB	24.22s
5	GyricAdam (gyric_adam)	1.6284	1.6763	5.3456	41.93%	39.80%	+0.1794	0.9776	11.10 dB	28.94s
6	CyclideAdam (cyclide_adam)	1.6440	1.6895	5.4169	41.43%	39.42%	+0.1596	0.9775	11.25 dB	25.22s
7	SymplecticAdam (symplectic_adam)	1.6534	1.7020	5.4851	41.04%	39.06%	+0.1485	0.9760	11.07 dB	29.75s
8	HuygensSymplecticAdam (huygens_symplectic_adam)	1.6531	1.7020	5.4852	40.87%	38.95%	+0.1491	0.9760	11.07 dB	29.74s

Key Geometric Findings on Rotational Text-to-Speech (GRTTTS)

• HessianPairAdam (Sophia-Torus) Leads Discrete RVQ Loss & Text-Conditioned Speech Discrimination: HessianPairAdam achieves #1 held-out RVQ validation loss (1.6574, val_ppl = 5.2456), #1 Codebook-0 primary acoustic accuracy (40.30%), and #1 text-conditioning margin (+0.1942 nats) over shuffled prompts.
• PolarNaturalAdam Achieves #1 Multi-Codebook RVQ Accuracy (42.37%): By scaling angular gradients by the Riemannian polar metric 1 / (r² + ε) alongside Cauchy–Riemann curvature, PolarNaturalAdam achieves the highest mean top-1 accuracy (42.37%) across all 4 RVQ speech codebooks ([40.24%, 40.41%, 43.48%, 45.34%]).
• PairAdam Peaks on Continuous Polar Spectrogram SNR (11.37 dB): On continuous GRTPolarVocoderHead waveform synthesis, PairAdam achieves 0.9788 frame polar cosine similarity and 11.37 dB spectral SNR in 26.17s.

---

Quickstart (transformers)

import torch
from transformers import AutoTokenizer, GRTTTSForConditionalGeneration

model_id = "MatterWaveAI/grt-30m-tts"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = GRTTTSForConditionalGeneration.from_pretrained(model_id, dtype=torch.float32)

prompt = "<|im_start|>user\nSpeak (alto, declarative): hello matter wave<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
outputs = model(**inputs)
print("Synthesized waveform shape:", outputs.speech_waveform.shape)

---

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

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

---

License

Apache 2.0

---

Citation

@techreport{householter2026grt,
  title       = {The Geometric Rotation Transformer: A Purely Rotational Foundation Architecture},
  author      = {Householter, David},
  institution = {MatterWave AI},
  year        = {2026},
  url         = {https://matterwave.ai}
}