Nikolajr93/hal-govornica-models

Whisper large-v3-turbo LiteRT INT8 export

This repository contains an independently produced LiteRT / TensorFlow Lite export derived from OpenAI Whisper large-v3-turbo. It is intended for on-device automatic speech recognition on Android using LiteRT.

This artifact is not an official OpenAI, Google, or Qualcomm release. Those organizations have not endorsed or certified this export.

Artifact

• File: whisper_large_v3_turbo_30s_kv448_i8.tflite
• Size: 827353216 bytes
• SHA-256: 888f4b53956204abf256753f467812a620c4e927bf5b23f99e8c14fb41184ae9
• Direct download: https://huggingface.co/Nikolajr93/hal-govornica-models/resolve/main/whisper_large_v3_turbo_30s_kv448_i8.tflite

After downloading, verify the file before use:

shasum -a 256 whisper_large_v3_turbo_30s_kv448_i8.tflite

The result must be:

888f4b53956204abf256753f467812a620c4e927bf5b23f99e8c14fb41184ae9

Export characteristics

• Base model: OpenAI Whisper large-v3-turbo
• Format: LiteRT / TensorFlow Lite (.tflite)
• Audio input window: 30 seconds
• Acoustic input: 128-bin log-Mel spectrogram
• KV cache: static
• KV / decoder context length: 448 tokens
• Quantization: INT8
• Signatures:
  • encode
  • cross_kv
  • prefill
  • decode (single token)

Intended use

The artifact is intended for local, on-device speech recognition on Android through LiteRT. Integrators are responsible for supplying compatible audio preprocessing, tokenizer data, decoding logic, and application-level testing. Model quality, latency, and hardware compatibility can vary by device and runtime version.

License and attribution

This downstream export is distributed under the MIT License. OpenAI Whisper is Copyright (c) 2022 OpenAI and is licensed under the MIT License. The upstream copyright and license notice are retained in LICENSE.

Upstream project: https://github.com/openai/whisper