Kokoro¶

Current status and review date¶

Decision snapshot

Product role: published co-resident TTS component in the retained 2026-07-28 RTX 5090 voice shape; not a claim about current live state.
Selected or best-qualified configuration: managed OpenAI-compatible Kokoro TTS service with a 2,048 MiB reservation.
Measured hardware: RTX 5090 on Primary Node, co-resident with Parakeet and the small Omni stack.
Evidence: functional and bounded capacity; one retained round trip measured 289.27 ms TTS and RTF 0.1006.
Decision: current in the dated optional co-resident voice evidence; not an LLM or router promotion.
Important limitation: exact model checkpoint revision is not retained, and the synthetic corpus is not primary-human TTS quality evidence.
Review dates: retained evidence through 2026-07-28; dossier-format review 2026-08-31.

Open the retained voice Compose configuration or jump to the decision, known limitations, or dated evidence.

Review narrative¶

2026-07-27 — Co-resident voice round trip¶

Kokoro served through the managed TTS endpoint as kokoro. In the retained small-Omni plus voice stack, the dedicated Parakeet/Kokoro round trip completed with 0.0 WER for the bounded utterance: 710.68 ms total, 421.41 ms STT, 289.27 ms TTS, and TTS RTF 0.1006. The measurement establishes a working co-resident TTS path; it is not a broad voice-quality study.

Outcome: retain Kokoro as the dated optional co-resident TTS component.

2026-07-28 — ASR corpus support¶

Kokoro generated the six bounded synthetic voice-agent phrases used alongside 24 human recordings in the ASR comparison. Those samples cover weather, timers, ZIP codes, dates/times, extensions, and correction/cancellation, but they are not primary-human quality evidence for Kokoro.

Older Mini-local research remains topology history and does not describe the retained Dark result. Historical broadcast-shape errors likewise do not invalidate this measured co-resident path.

Outcome: synthetic-corpus use is supporting provenance, not a TTS quality qualification or a new deployment decision.

Immutable identity¶

Served name: kokoro.
Service: managed OpenAI-compatible TTS endpoint.
Observed image identity: local image ID sha256:81a6937f1ea24cd4a729de517385074cb4ee113ed9ba70bde0557ac912451712 in the retained 2026-07-28 runtime-state artifact.
Checkpoint repository and revision: Not retained. Do not infer a model commit from the served alias.
Registry image digest: Not retained. The public Compose reference ghcr.io/remsky/kokoro-fastapi-gpu:latest-cu128 is mutable and is not an immutable reconstruction by itself.

Tested hardware and topology¶

Measured: RTX 5090 on Primary Node.
Co-resident: Parakeet STT and the small Omni stack.
Reservation: 2,048 MiB for TTS.
Historical-only topology: Mini-local experiments are not the retained reference path described here.

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

Retained Dark TTS service¶

Runtime: Kokoro OpenAI-compatible TTS service.
Model alias: kokoro.
KV and text context: Not applicable to this TTS endpoint.
Concurrency contract: Not recorded in the retained dossier evidence.
Recipe: voice Compose service and managed serve manifest.

Evidence by measurement class¶

Co-resident TTS path¶

Status: functional, bounded capacity.
Measured: 289.27 ms TTS, RTF 0.1006, and 138,082 output bytes inside a 710.68 ms dedicated voice round trip.
Limit: one bounded round trip is not a distribution or a primary-human listening-quality result.
Evidence: voice-stack finding and raw round-trip artifact.

Synthetic corpus contribution¶

Status: corpus provenance support, not a separate model qualification.
Measured: six generated agent-style phrases were included in the ASR corpus; synthetic results were reported separately from the 24 human cases.
Evidence: ASR qualification and retained runtime state.

Decision and promotion state¶

Retained dated role¶

Kokoro TTS: current in the optional co-resident voice shape published through 2026-07-28. This is not a claim about live operator state.

Promotion boundary¶

The TTS result is not an LLM or router promotion and does not authorize a topology change.

Failures and gotchas¶

Evidence and interpretation limits¶

Checkpoint identity: exact model repository and revision are not retained, so an exact model rerun cannot be claimed.
Quality boundary: synthetic-agent WER must remain separate from primary-human evidence; the retained TTS timing is not a listening test.

Runtime and topology limits¶

Mutable image reference: latest-cu128 can drift; the retained local image ID documents the old runtime but is not a portable registry digest.
Historical errors: older broadcast-shape errors and Mini-local research do not describe the retained Dark result.