--- license: mit tags: - ava - voiceos - whisper - speech-recognition - qnn - qualcomm - on-device base_model: - openai/whisper-tiny - openai/whisper-base - openai/whisper-small --- # AVA Whisper Models (QNN) On-device speech-to-text models for VoiceOS, compiled for Qualcomm QNN. ## Models | AVA ID | Source | Size | Chipsets | Priority | |--------|--------|------|----------|----------| | AVA-WHISPER-TINY | openai/whisper-tiny | ~50 MB | QCS8550, SD8G3, SD8E | P0 | | AVA-WHISPER-BASE | openai/whisper-base | ~150 MB | QCS8550, SD8G3, SD8E | P0 | | AVA-WHISPER-SMALL | openai/whisper-small | ~500 MB | QCS8550, SD8G3, SD8E | P1 | ## Target Chipsets | Chipset | Device | Proxy | |---------|--------|-------| | QCS8550 | Vuzix LX1 (QCS4490 compatible) | qcs8550_proxy | | Snapdragon 8 Gen 3 | Flagship phones 2024-2025 | snapdragon_8gen3 | | Snapdragon 8 Elite | Flagship phones 2025-2026 | snapdragon_8_elite_gen5 | ## Directory Structure ``` raw/ # Source ONNX from Qualcomm tiny/ encoder.onnx decoder.onnx base/ small/ production/ # AON-encrypted per chipset whisper-tiny-qnn-qcs8550.aon.zip whisper-tiny-qnn-sd8g3.aon.zip whisper-base-qnn-qcs8550.aon.zip ``` ## License Raw models: MIT (OpenAI). QNN compilation: Qualcomm. AON-encrypted: Proprietary (Intelligent Devices LLC).