Parakeet TDT 0.6B v3¶

Current status and review date¶

Decision snapshot

Product role: Published routed-STT reference baseline in the retained July 2026 evidence.
Selected or best-qualified configuration: parakeet.cpp-server CUDA image at the managed STT endpoint, tested sequentially and at concurrency four with normalized 16-kHz mono WAV input.
Measured hardware: One RTX 5090 on Primary Node; the RTX PRO 6000 was protected and left running during the final corpus campaign.
Evidence: 3.343% primary-human micro-WER, 177.87 ms sequential p95, 240.43 ms concurrency-four p95, and zero final failures or repetitions.
Decision: Retain as the published baseline; neither July challenger result authorized a route change.
Important limitation: The model repository is recorded, but its exact checkpoint commit is not retained.
Review dates: Retained evidence cutoff: 2026-07-28. Dossier-format review: 2026-08-31.

Review narrative¶

2026-07-08 — Initial routed-STT baseline¶

The earlier STT benchmark established Parakeet as the routed reference used by the later comparison. That record is historical evidence, not a statement about unverified live state after the retained evidence cutoff.

2026-07-28 — Multi-sample corpus qualification¶

The final normalized WAV corpus measured 3.343% primary-human micro-WER with 177.87 ms sequential p95 and 240.43 ms concurrency-four p95. It completed with zero final failures or repetitions while the RTX PRO 6000 workload remained protected. Neither challenger result authorized a route change.

Immutable identity¶

Model¶

Repository: nvidia/parakeet-tdt-0.6b-v3.
Checkpoint commit: Not retained.

Runtime¶

Runtime: parakeet.cpp-server CUDA image.
Retained image ID: sha256:7a08005a8a26dd8fe3709f1df3d8dc44be7afef93a27e0830916cc2f54d0304e.

Tested hardware and topology¶

Final corpus lane¶

Host label: Primary Node.
Measured device: one RTX 5090.
Protected co-resident device: one RTX PRO 6000, left running and not used for the measured STT candidate.

Other accelerator products were not tested by this retained campaign.

Engine, quantization, KV, context, and concurrency recipe¶

Qualified corpus recipe¶

Server: parakeet.cpp-server CUDA image at the managed STT endpoint.
Audio preparation: normalized 16-kHz mono WAV.
Schedules: sequential and concurrency four.
Quantization, KV cache, and language-model context: Not applicable to this speech-recognition recipe.

The public STT experiment overlay records the reconstructable service configuration.

Evidence by measurement class¶

Quality¶

Primary-human micro-WER: 3.343%.
Final repetitions: zero.

Capacity and latency¶

Sequential p95: 177.87 ms.
Concurrency-four p95: 240.43 ms.
Final request failures: zero.

Evidence¶

The dated finding links the retained corpus artifacts and result summaries. The exact checkpoint commit remains not retained.

Decision and promotion state¶

Retained baseline¶

Published routed-STT reference in the July 2026 evidence.
Neither challenger result authorized a route change.

Current-state boundary¶

This dossier does not claim that the same route is live after the retained evidence cutoff.

Failures and gotchas¶

Input-format failure¶

Initial FLAC requests failed closed and remain retained as incomplete artifacts. They are not counted in the final corpus result.

Evidence limitation¶

The exact checkpoint commit is not retained. Reproduction must therefore distinguish the pinned runtime image from the unpinned model snapshot.