Sorika Labs · Architecture MatrixUpdated Sep 2026
Foundation Suite
Architecture Matrix.
Specifications, empirical benchmarks, and deployment profiles across Sorika Labs' compact foundation models — architected by Darsh Yadav.
Speech AI FoundationSuzune S1 Voice
PRODUCTIONSpeech Recognition AIVaani 1 Lite ASR
PRODUCTIONEdge Reasoning & Spatial ActionSwen Series & Fast Thinking
R&DPRODUCTION
Suzune S1 Voice
Speech AI Foundation80M non-autoregressive neural TTS with sub-20ms streaming latency and 24kHz HD broadcast audio.
Parameters80M
Latency< 20 ms TTFA
VRAM~320 MB O(1)
StatusLive Production v2.5
PRODUCTION
Vaani 1 Lite ASR
Speech Recognition AISub-100M streaming multilingual Conformer architecture for low-resource Indic dialects and real-time transcription.
ParametersSub-100M
Latency< 100 ms Streaming
VRAM~380 MB O(1)
StatusLive Production v2.5
R&D
Swen Series & Fast Thinking
Edge Reasoning & Spatial ActionNext-gen spatial controller & metacognitive reasoning models with hybrid conv-attention backbones.
ParametersSub-100M → 1.2B
Latency< 15 ms TTFT
VRAMBounded O(1)
StatusActive R&D — Next 2 Months
Detailed Specification Matrix
Core Architecture
Primary ModalityInput / output data domain
Suzune S1Text → Speech (24kHz)
Vaani 1 LiteAudio → Text (Streaming ASR)
Swen SeriesSpatial Embeddings & Logic Traces
Model FootprintTotal trainable weights
Suzune S180 Million Parameters
Vaani 1 LiteSub-100M Parameters
Swen SeriesSub-100M (Nano) → 1.2B
Neural FormulationUnderlying algorithmic paradigm
Suzune S112-L PL-BERT + Fast iSTFTNet Vocoder
Vaani 1 LiteMulti-scale Conformer + Neural Transducer
Swen SeriesHybrid Causal Conv + Linear GQA
Memory ScalingVRAM behavior across long contexts
Suzune S1Strictly Constant O(1) ~320 MB
Vaani 1 LiteConstant O(1) ~380 MB
Swen SeriesBounded Constant O(1) Cache
Performance Telemetry
Execution LatencyTime to first token / audio
Suzune S1< 20ms TTFA (GPU)
Vaani 1 Lite< 100ms Chunk Latency
Swen Series< 15ms TTFT (edge target)
Key BenchmarkStandardized evaluation metric
Suzune S1MOS 4.38 ± 0.06
Vaani 1 Lite7.8% Indic WER
Swen SeriesDeliberative CoT Scratchpads
Inference HardwareOptimal execution environment
Suzune S1Consumer CPUs, Mobile, GPUs
Vaani 1 LiteARM/x86 CPUs, Edge, GPUs
Swen SeriesLaptops, Embedded, Local GPUs
Language SupportNative multilingual breadth
Suzune S1English & Hindi (Expressive)
Vaani 1 Lite22 Scheduled Indic Languages + Code-Switching
Swen SeriesMultilingual Token Mixing
Deployment & Availability
Lifecycle StateProduction readiness
Suzune S1Live Production v2.5
Vaani 1 LiteLive Production v2.5
Swen SeriesActive R&D — Next 2 Months
Interactive StudioWeb-based exploration
Suzune S1Live — Voice Studio
Vaani 1 LiteLive — ASR Pipeline
Swen SeriesPreview Roadmaps
Air-Gapped DeploymentOn-premise zero-retention
Suzune S1Full On-Prem / Local
Vaani 1 LiteFull On-Prem / Local Edge
Swen SeriesLocal Edge Target