oruk/orukeet

1# Orukeet for transcribe.cpp2 3This Q8_0 export uses transcribe.cpp's existing Parakeet TDT v3 implementation. The 12,288 fitted Gabor kernels are materialized as ordinary convolution weights. There are no additional runtime operators or dependencies.4 5- File: `orukeet-transcribe-cpp-Q8_0.gguf` (739,508,608 bytes)6- SHA-256: `cad2f52ac91cad829279422301989687c2cf02e19157352ed25ea501b90dbb7e`7- Source: Orukeet r3, SHA-256 `031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56`8- Weights: [CC BY-SA 4.0](../LICENSE-WEIGHTS). Orukeet is an adaptation of NVIDIA Parakeet TDT v3; retain [attribution](../NOTICE.md).9 10The root-level filename follows Handy's existing cache and delete layout. The earlier `transcribe-cpp/orukeet-Q8_0.gguf` path remains available and contains identical bytes.11 12## Run13 14Build [transcribe.cpp](https://github.com/cjpais/transcribe.cpp/tree/585b98f7e66777d16f2da734ceedaa7398060fa7) normally, then run:15 16```sh17./build/bin/transcribe-cli --model orukeet-transcribe-cpp-Q8_0.gguf recording.wav18```19 20Use 16 kHz mono audio. This is offline transcription in 25 languages, with language detection and token timestamps. Translation and streaming recognition are not supported. Use this layout with transcribe.cpp; the root-level `orukeet-v0.1.0-q8.gguf` is for NeMo-Speech.cpp.21 22## Reproduce the export23 24At the converter commit in [manifest.json](manifest.json), apply [convert-orukeet.patch](convert-orukeet.patch), then:25 26```sh27uv run --no-project --python 3.11 --with torch --with omegaconf --with sentencepiece --with gguf scripts/convert-parakeet.py orukeet-v0.1.0.nemo orukeet-F32.gguf --repo-id oruk/orukeet28./build/bin/transcribe-quantize orukeet-F32.gguf orukeet-Q8_0.gguf --quant Q8_029```30 31## Validation32 33The exact transcribe-cpp 0.2.0 Rust dependency pinned by Handy loads and transcribes this file on CPU and Apple Metal. Each device passed 120 multilingual clips, repeated decoding, 0.1/1/5-second silence, cancellation, and successful session reuse after cancellation. The supporting C++ build passed all 38 tests.34 35On the fixed six-language FLEURS validation sample (120 clips, 2,433 reference words), Q8 Metal and the NeMo FP32 source both score **5.34% pooled WER**; Q8 CPU scores **5.30%**. Metal matches the source text exactly on 108 clips and after word normalization on 112. The NeMo reference uses transcribe.cpp's pinned NeMo 2.8.0rc0 environment. [All paired language scores and edit counts](validation.json).36 37| Q8 Metal WER | Parakeet TDT v3 | Orukeet |38| --- | ---: | ---: |39| English | 4.19% | 3.40% |40| German | 3.56% | 4.07% |41| Spanish | 3.41% | 3.01% |42| French | 4.07% | 4.50% |43| Russian | 7.77% | 7.51% |44| Ukrainian | 9.40% | 11.60% |45| Pooled | 5.14% | 5.34% |46 47Standard source/export tensor comparisons pass the existing upstream Parakeet tolerances and the JFK reference transcript matches exactly. Four additional sub-block probes exceed generic strict tolerances, with the largest differences at sequence boundaries. The runtime and NeMo use different valid-length padding paths; no runtime code or numerical tolerance was changed for this export.48 49These are integration checks for this export and runtime, separate from the report's NeMo benchmark. No Windows/Vulkan or comparative speed claim is made here.50