ASR phase 1: a sovereign Whisper-tiny path landed
Ran Tao (Octoryn Research)
Abstract
Engineering milestone: a sovereign Whisper-tiny ASR path (mel spectrogram to encoder to decoder to greedy decode) was integrated into two engines and asserted by a green test suite (no failures) against a bundled golden reference. This is golden-parity correctness-smoke, NOT a WER/accuracy or throughput result; a release build is required for practical run time. Scope is limited to Whisper-tiny greedy decoding only — larger models, quantization, beam search, and streaming are untested.
Contribution
Claimed. A sovereign Whisper-tiny automatic-speech-recognition path — mel spectrogram -> encoder -> decoder -> greedy decode — has been integrated into two engines (an audio engine and a speech-provider engine). The integration's correctness is asserted by a green test suite (no failures) against a bundled golden reference. This is an engineering integration and correctness-smoke result: it establishes the ASR seam (the engine boundary and the golden-harness scaffold) and confirms the phase-1 end-to-end path produces output matching a pre-recorded golden.
NOT claimed. This note does not claim any accuracy/WER measurement on real audio benchmarks — there is no transcription-quality evaluation here, only golden parity. It does not claim throughput or latency numbers (the only performance fact is qualitative: the naive forward is impractically slow in a debug build and must be run in a release build). It does not claim coverage of larger Whisper variants, quantization, or streaming. It is not a novel algorithm; it is a sovereign re-implementation landing a known model on a green test seam.
Methods
| Field | Value |
|---|---|
| Model | Whisper-tiny |
| Path | mel -> encoder -> decoder -> greedy decode |
| Integration targets | audio engine, speech-provider engine |
| Oracle / baseline | bundled golden reference (present); external oracle and version not recorded in source notes |
| Procedure | run the engine test suites; assert against the bundled golden |
| Build / run requirement | release build mandatory (debug forward impractically slow) |
| Hardware | Apple Silicon Mac (M-series) |
| Sovereignty | no third-party ML framework (vendor-optimized math primitives only) |
| Precision | not recorded in source notes |
| Metric + gate | test pass/fail (golden parity); WER not measured |
| Stimulus / inputs | bundled golden-harness inputs; exact audio clips not recorded in source notes |
Results
| Measurement | Value |
|---|---|
| Test suite | green (no failures) |
| Release-configuration subset | runs and passes |
| Golden | present |
Threats to validity / Limitations
- Correctness smoke, not a benchmark. The evidence is golden-parity over a bundled harness, not an accuracy evaluation. No WER, CER, or any transcription-quality metric is reported. Parity with a golden does not establish recognition quality on real or diverse audio.
- Sample scope. The number, length, and diversity of the golden audio inputs are not recorded in the source; the test count reflects test cases, not audio samples, and must not be read as benchmark-scale coverage.
- No performance data. The only performance statement is qualitative ("naive forward impractically slow in a debug build"). There are no latency, throughput, or real-time-factor numbers, and therefore no variance or single-run caveats to report — because no number exists.
- Scope bounds. Results apply only to Whisper-tiny via the phase-1 greedy path in a release build. Larger models, quantization, beam search, and streaming are explicitly out of scope and untested.
- Oracle/precision gaps. The external oracle (if any) and its version, and the numerical precision of the path, are not recorded in the source notes; golden parity is only as trustworthy as the (unvalidated-here) golden's own provenance.
- Evidence level. Marked "experimental"; given the absence of any quantitative quality or performance metric, even that label leans on the green-test signal alone — a skeptical reader should treat this as an integration milestone, not a validated result.
Reproducibility / Provenance
| Field | Value |
|---|---|
| Run requirement | release build mandatory (debug forward impractically slow) |
| Engines | audio engine, speech-provider engine |
| Hardware | Apple Silicon Mac (M-series) |
| Sovereignty | no third-party ML framework (vendor-optimized math primitives only) |
| Commit(s) | not recorded in source notes |
| Version pins | not recorded in source notes |
| Golden | bundled (present); origin/validation of the golden not recorded in source notes |
Caveat: the golden's own correctness is assumed, not independently validated in this note — a reproducer should first establish the golden's provenance before treating parity as a correctness proof.
Claim boundary
The author's explicit scope — what this work does and does not establish — carried over from the Octoryn Research publishing model.
Proves
- A sovereign Whisper-tiny mel-to-encoder-to-decoder-to-greedy path was integrated into two engines with a green test suite (no failures) and a bundled golden reference present.
Does not prove
- Accuracy/WER or any transcription-quality metric on real audio benchmarks (golden parity is not recognition quality).
- Larger Whisper models, quantization, beam search, or streaming.
- Throughput, latency, or real-time-factor (only a qualitative 'release build required' note exists).
- Benchmark-scale coverage — test count is not audio-sample count; input diversity is not recorded.
Applies when
- The phase-1 Whisper-tiny greedy path with the bundled golden, run in a release build configuration.
Does not apply when
- WER/accuracy, larger-model, quantization, beam-search, or streaming claims.
- Any performance/throughput claim, or generalization beyond the bundled golden inputs.
Authors
- Ran Tao — Investigation, Writing
Cite this
Citation
Tao, R., Octoryn Research. (2026). ASR phase 1: a sovereign Whisper-tiny path landed (RN-2026-0007). Octopus Research Institute.
BibTeX
@techreport{orirn20260007,
title = {ASR phase 1: a sovereign Whisper-tiny path landed},
author = {Tao, Ran and {Octoryn Research}},
institution = {Octopus Research Institute},
year = {2026},
note = {Permanent ID RN-2026-0007. Not peer reviewed.}
}Disclosures
- Funding
- Hardware and infrastructure provided by Octoryn / Octopus Core Pty Ltd.
- Conflicts of interest
- Octoryn ships commercial inference and governance tooling; findings are reported independently.