Śabda-ALM · a speech-centred language model
Bhartṛhari's śabdādvaita: language is not mere description but the ultimate itself. We built a speech model on this principle — the four vāk (speech levels) as the architecture. Result: 0.0392 cer_norm (4.8× better than open ASRs) after honestly correcting two harness bugs. The Sphoṭa-Lens shows you layer 13, where sound becomes meaning.
The thesis
Bhartṛhari's Vākyapadīya describes how meaning bursts from continuous sound — the sphoṭa. The four levels of speech — parā, paśyantī, madhyamā, vaikharī — form a gradient from unmanifest source to uttered surface. A text language model is frozen at the final layer (vaikharī) with no acoustic substrate and no emergence gradient. An audio language model walks the whole path. Therefore: sphoṭa is computationally realizable through an ALM, not a text-only LLM.
The source — the model's Sanskrit language prior, whole and silent.
The flash of meaning. Layer where sound becomes sense — measured at layer 13.
Inner articulation — the Sphoṭa Projector that turns continuous frames into tokens.
The uttered word — audio in, and response out. The acoustic interface.
The architecture
NVIDIA's Parakeet encoder (frozen, 0.6B) hears the audio. A trainable Sphoṭa Projector turns acoustic frames into tokens the Sanskrit core understands. The core (Nemotron-H/Mamba-2, from-scratch Sanskrit byte-core, 200M→1.13B) reasons. Neural TTS turns the answer back into sound. Only the projector and a few adapter weights train.
The Sphoṭa-Lens
At every layer we ask: how well can the sentence's meaning (its verb, the core concept) be read from the audio positions alone? The answer rises to a peak at layer 13, then declines toward the output. That peak is the computational sphoṭa: where the flash of meaning happens. A causal test (ablating each layer) agrees: correlational peak (layer 13) and causal peak (layer 14) agree within ±1 layer. Meaning is decodable from sound at 11× chance (0.263 vs 0.022 random).
◈ Layer 13 (200M model): correlational + causal peaks agree. The paśyantī → vaikharī gradient made measurable.
The corrected benchmark
Earlier result claimed 28× superiority. It was wrong. Bug 1: Qwen's audio kwarg was silently ignored (transformers 4.57). Bug 2: our specialist got a gold-length oracle. We report both bugs, fix both, and re-run. Fair protocol: identical frozen 58-clip val, greedy decode, 64-byte budget, no oracle, scheme-neutral metric (cer_norm). This correction is the paper's honesty exemplar.
| Model · fair eval | Type | Params | cer_norm ↓ | |
|---|---|---|---|---|
| 1 | Śabda-ALM 1.13B+LoRA XL — ours | specialist | 1.13B+8.4M | 0.0392 |
| 2 | Voxtral-Mini-3B-2507 (Mistral) | generalist | 3B | 0.1866 |
| 3 | Qwen2.5-Omni-3B Thinker (Alibaba) | generalist | 3B | 0.2133 |
| 4 | Qwen2-Audio-7B-Instruct (Alibaba) | generalist | 7B | 0.4305 |
4.8× lower than Voxtral-Mini (0.1866 ÷ 0.0392). Trained on 9,959 clips (5.77h native-Sanskrit TTS), 3 epochs, best checkpoint epoch 1 (no overfit). Data-scaling curve: 1.82 → 1.03 → 0.46 → 0.076 → 0.039 (scale + curriculum + LoRA all matter).
Injecting a concept direction into the workspace band yields uptake 0.767 (5.11× random control). But entropy collapses to −1.436 nats (budget 0.5 nats breached). Caveat: loading into the workspace ≠ faithful utterance.
Pramāṇa-style legality constraint (answer grounded in transcript) fires in 94.69% of cases but improves accuracy 0 → 0. Finding: transcription quality is the bottleneck, not the kāraka decoder.
Honesty boundary
Full transparency
| Claim | Evidence File | Gate |
|---|---|---|
| Sphoṭa realizable via ALM, not text | THESIS-sabdadvaita.md | Architectural |
| cer_norm 0.0392 on fair benchmark | data/benchmark/alm_vs_alm.json | AA (apples-to-apples) |
| 4.8× lower than Voxtral (0.1866) | alm_vs_alm.json · arithmetic | AA |
| Bug 1: Qwen audio kwarg ignored | research/alm-vs-alm.md §1 | AA |
| Bug 2: specialist oracle removed | research/alm-vs-alm.md §2 | AA |
| Data curve 1.82→0.039 | research/alm-vs-alm.md · alm metrics | AA |
| Sphoṭa layer 13: 0.263 decodability | data/sphota_lens/emergence_report.json | SU (steering) |
| Steering uptake 5.11× (0.7667 vs 0.15) | data/alm/p3su_results.json | SU |
| Entropy delta −1.436 (H-SU3 failed) | data/alm/p3su_results.json | SU |
| Nyāya fire-rate 0.9469, Δ accuracy 0 | data/alm/p3ng_results.json | NG (nyāya) |
Reproducibility: Pre-registration at research/prereg/P3-battery-prereg.md (locked 2026-07-18). Run all gates: python gates/check.py AA, python gates/check.py SU, python gates/check.py NG.