ahhhhhh3/parakeet-ultra-vulkan

Parakeet Ultra encoder for Android Vulkan

An FP16 Vulkan export of the encoder and encoder projector from Moondream Parakeet Ultra, based on NVIDIA Parakeet TDT 0.6B v3. Converted by ahhhhhh3 for local Android transcription. This is an encoder component, not a standalone speech recognizer or an official Moondream/NVIDIA release.

Files and runtime

• encoder-vulkan-30s.pte: 1,219,557,506 bytes; SHA-256 76bc8de211184fbbc4388c45732b873be5b8250571ede08d49b6dcc4410ac506.
• ExecuTorch Android Vulkan 1.5.1; application admission requires Vulkan 1.1.
• Input: float32 normalized log-mel features [1, 128, 3000].
• Output: float32 projected encoder features [1, 640, 375].
• Fixed 30-second chunks, 16 kHz mono audio. Pad the final chunk with silence before feature extraction. This file does not accept audio directly.
• Uses FP16 Vulkan computation internally, exported with VulkanPartitioner({'force_fp16': True}) and to_edge_transform_and_lower.
• Source weights: moondream/parakeet-ultra revision 73175eb7aeb0d82f1e2a6b53b3aabc10a90bcd0b.

The tested application uses NeMo-compatible 128-bin log-mel features with per-feature population normalization, and the int8 decoder, joiner and tokens from mldecode/parakeet-ultra-onnx-int8. Those separate files and a compatible TDT decoder are required to produce text.

Validation and limitations

Tested in a local Android app on Pixel 7a only. A 23:59 synthetic recording completed in 230.620 seconds including MP3 conversion and model loading, versus 384.942 seconds in an earlier int8 CPU app run. These were not randomized trials. The phone was locked for most of the GPU run and woke near the end.

The GPU transcript had 74 normalized word edits over 4,413 reference words (1.68%), versus 52 edits (1.18%) in the earlier int8 CPU comparison. The beginning and end were preserved. This is a small accuracy cost on one synthetic fixture, not evidence of general accuracy parity. Runtime memory was approximately 1.6 GiB for the GPU worker; total app memory is higher.

Other GPUs, Android vendors, languages, thermal behavior and battery consumption are unverified. An app should validate output and speed on the actual device and retain a CPU path if GPU validation or execution fails. The tested app isolates GPU execution in a separate process so native failures can recover.

Attribution and license

Model weights are distributed under Creative Commons Attribution 4.0 International, as declared by the upstream Moondream model. Credit Moondream for Parakeet Ultra and NVIDIA for the underlying Parakeet model. This conversion changes the execution format, uses FP16 Vulkan arithmetic and includes only the encoder and projector; it is not a new training run. Retain attribution, link the license, and identify modifications when redistributing. Runtime software has its own separate licenses.

Only the converted model and documentation are hosted here. Inference happens on the user's device; this repository is not a hosted transcription service.