--- license: other license_name: proprietary tags: - tali - taliOS - on-device - automatic-speech-recognition --- # Tali STT **Tali STT** is the on-device speech-to-text engine for **TaliOS**, compiled for mobile NPU (Neural Processing Unit) acceleration. ## What it does Tali STT is the first model in the TaliOS voice pipeline. It listens to microphone audio and converts spoken words into text — continuously, at low latency, on the user's device. Audio never leaves the hardware: there is no cloud transcription path in the production runtime. ``` User speech ──► Tali STT ──► text ──► Tali NLU / Storm / Nautilus (audio) (this) (UTF-8) (downstream language models) ``` The text Tali STT produces is the input to every downstream TaliOS language model. Quality and latency of the rest of the voice pipeline are gated by this model. ## Why NPU Mobile NPUs (the Hexagon DSP/AI co-processor on Snapdragon, equivalents on other SoCs) run audio inference at a fraction of the power cost of the main CPU. Tali STT models in this repository are pre-compiled per chipset family so the runtime can load them directly — no on-device compilation step at install time. ## Variants | Tali ID | Size | Target Chipsets | Priority | |-----------------|----------|--------------------------------|----------| | TALI-STT-TINY | ~50 MB | QCS8550, SD8G3, SD8E | P0 | | TALI-STT-BASE | ~150 MB | QCS8550, SD8G3, SD8E | P0 | | TALI-STT-SMALL | ~500 MB | QCS8550, SD8G3, SD8E | P1 | The TaliOS runtime selects the largest variant the device can support, falling back to a smaller one if NPU memory is constrained. ## Target Chipsets | Chipset | Device class | |---------------------|-------------------------------------------| | QCS8550 | Smart-glass class (Vuzix LX1 / QCS4490 compatible) | | Snapdragon 8 Gen 3 | Flagship phones 2024-2025 | | Snapdragon 8 Elite | Flagship phones 2025-2026 | ## File format Shipped weights are wrapped in **AON** — Tali's encrypted, signed asset container. The `.aon` extension is the only format the OS or external tooling sees. AON files are decrypted in-process by the TaliOS runtime; they cannot be loaded by third-party inference engines. ## License Proprietary — Intelligent Devices LLC.
