शब्द

Śabda-ALM · a speech-centred language model

Sound becomes meaning. Text cannot.

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.

0.0392cer_norm on fair benchmark
4.8×better than Voxtral-Mini
Layer 13where sound becomes meaning
100% honestbugs corrected & reported

The thesis

Sphoṭa is realizable through audio, not through text.

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.

परा

Parā

The source — the model's Sanskrit language prior, whole and silent.

पश्यन्ती

Paśyantī · Sphoṭa

The flash of meaning. Layer where sound becomes sense — measured at layer 13.

मध्यमा

Madhyamā

Inner articulation — the Sphoṭa Projector that turns continuous frames into tokens.

वैखरी

Vaikharī

The uttered word — audio in, and response out. The acoustic interface.

The architecture

Four levels of speech, as a neural pipeline.

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.

vaikharī in
Audio
continuous sound
dhvani
Parakeet encoder
frozen · 0.6B
madhyamā
Sphoṭa Projector
trained
paśyantī
Sanskrit byte-core
200M→1.13B · frozen then LoRA
vaikharī out
Text → TTS
spoken answer

The Sphoṭa-Lens

We measured where sound becomes meaning.

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).

meaning-decodability from sound · by layerchance = 0.022
L0 L24

◈ Layer 13 (200M model): correlational + causal peaks agree. The paśyantī → vaikharī gradient made measurable.

The corrected benchmark

Specialist leads after correcting two harness bugs.

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 evalTypeParamscer_norm ↓
1Śabda-ALM 1.13B+LoRA XL — oursspecialist1.13B+8.4M0.0392
2Voxtral-Mini-3B-2507 (Mistral)generalist3B0.1866
3Qwen2.5-Omni-3B Thinker (Alibaba)generalist3B0.2133
4Qwen2-Audio-7B-Instruct (Alibaba)generalist7B0.4305

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).

Steering works — with caveats

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.

Nyāya guardrail honest negative

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

What we do NOT claim.

Full transparency

Every claim traced to its gate.

ClaimEvidence FileGate
Sphoṭa realizable via ALM, not textTHESIS-sabdadvaita.mdArchitectural
cer_norm 0.0392 on fair benchmarkdata/benchmark/alm_vs_alm.jsonAA (apples-to-apples)
4.8× lower than Voxtral (0.1866)alm_vs_alm.json · arithmeticAA
Bug 1: Qwen audio kwarg ignoredresearch/alm-vs-alm.md §1AA
Bug 2: specialist oracle removedresearch/alm-vs-alm.md §2AA
Data curve 1.82→0.039research/alm-vs-alm.md · alm metricsAA
Sphoṭa layer 13: 0.263 decodabilitydata/sphota_lens/emergence_report.jsonSU (steering)
Steering uptake 5.11× (0.7667 vs 0.15)data/alm/p3su_results.jsonSU
Entropy delta −1.436 (H-SU3 failed)data/alm/p3su_results.jsonSU
Nyāya fire-rate 0.9469, Δ accuracy 0data/alm/p3ng_results.jsonNG (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.