Backfill speech doctrine and narrow Phase 6R-F
This commit is contained in:
axiomlogicnexus
parent
b578342ee5
commit
2a1c24ebe5
6 changed files with 285 additions and 22 deletions
|
|
@ -124,21 +124,28 @@ Status update on `2026-05-21`:
|
|||
- `vivaansinghvi07/rubix-cube-solver` remains partially incorporated
|
||||
- the bounded permissive `Phase 6R-D` `roice3/MagicTile` tiling topology and geometry-family
|
||||
packet is now landed in current code
|
||||
- `roice3/MagicTile` remains partially incorporated; the current next bounded move is a
|
||||
source-backed `Phase 6R-E` speech-input / voice sidecar preparation/control pass, starting with
|
||||
`ggml-org/whisper.cpp`, not a new restrictive packet by default
|
||||
- `roice3/MagicTile` remains partially incorporated
|
||||
- the bounded permissive `Phase 6R-E` `ggml-org/whisper.cpp` speech transcript session packet is
|
||||
now landed in current code
|
||||
- `ggml-org/whisper.cpp` remains partially incorporated
|
||||
- the provider-family backfill now keeps the top-level speech/provider contract explicitly
|
||||
first-party and provider-neutral
|
||||
- the generic source-backed `Phase 6R-F` `SYSTRAN/faster-whisper` control pass is now consumed
|
||||
- the current next bounded move is the bounded permissive `Phase 6R-F` `SYSTRAN/faster-whisper`
|
||||
implementation packet for the Python transcription-service orchestration profile, not a new
|
||||
restrictive packet by default
|
||||
- use the repo-row README census, the portfolio standing refresh backfill, and the `2R-A`
|
||||
ownership contract for the current queue after that correction
|
||||
|
||||
This file now also preserves the current truth that future models must not lose:
|
||||
|
||||
- current curated HyperTwist shallow-eval set: `75` repos
|
||||
- currently verified live/implemented in checked `UnrealHyperTwist` surfaces: `28`
|
||||
- permissive live lanes: `17`
|
||||
- currently verified live/implemented in checked `UnrealHyperTwist` surfaces: `29`
|
||||
- permissive live lanes: `18`
|
||||
- boundary-sensitive live lanes: `6`
|
||||
- restrictive live lanes: `5`
|
||||
|
||||
The seventeen permissive live lanes are:
|
||||
The eighteen permissive live lanes are:
|
||||
|
||||
- `Aarav2709/KubeTimr`
|
||||
- `HactarCE/Hyperspeedcube`
|
||||
|
|
@ -157,6 +164,7 @@ The seventeen permissive live lanes are:
|
|||
- `poliva/cubedex`
|
||||
- `newyork-anthonyng/rubiks-cross-trainer`
|
||||
- `roice3/MagicTile`
|
||||
- `ggml-org/whisper.cpp`
|
||||
|
||||
The six boundary-sensitive live lanes are:
|
||||
|
||||
|
|
@ -227,8 +235,8 @@ Current practical interpretation:
|
|||
- the landed `PostHog/posthog` boundary-sensitive widening packet now lives in `docs/HYPERTWIST_PHASE_4R_PACKET_4R_D_POSTHOG_CONTROL_PLANE_IMPLEMENTATION_2026-05-13.md`
|
||||
- the landed `screenpipe/screenpipe` boundary-sensitive widening packet now lives in `docs/HYPERTWIST_PHASE_4R_PACKET_4R_E_SCREENPIPE_CAPTURE_HISTORY_IMPLEMENTATION_2026-05-13.md`
|
||||
- the landed `remotion-dev/remotion` boundary-sensitive widening packet now lives in `docs/HYPERTWIST_PHASE_4R_PACKET_4R_F_REMOTION_MEDIA_EXPORT_IMPLEMENTATION_2026-05-13.md`
|
||||
- the next bounded move is a source-backed `Phase 6R-E` speech-input / voice sidecar
|
||||
preparation/control pass, starting with `ggml-org/whisper.cpp`
|
||||
- the next bounded move is the bounded permissive `Phase 6R-F` `SYSTRAN/faster-whisper`
|
||||
implementation packet for the Python transcription-service orchestration profile
|
||||
|
||||
Companion docs:
|
||||
|
||||
|
|
@ -1422,6 +1430,8 @@ Approved working posture:
|
|||
- keep it bounded behind a speech-input seam
|
||||
- prefer command, dictation, and constrained coach-interaction use over broad “voice assistant platform” scope
|
||||
- the first bounded permissive implementation slice is now landed as a speech transcript session boundary above the existing companion listening lifecycle
|
||||
- the top-level provider/session contract remains first-party HyperTwist-owned and
|
||||
provider-neutral; this row does not own that lane
|
||||
- keep model files, downloadable payloads, and future voice-asset review separate from the code-license judgment
|
||||
- treat `SYSTRAN/faster-whisper` as the next complementary queue head for Python orchestration rather than widening `whisper.cpp` into the runtime core
|
||||
|
||||
|
|
@ -1431,6 +1441,7 @@ Decision date:
|
|||
|
||||
- `2026-04-24`
|
||||
- refreshed on `2026-05-13`
|
||||
- refreshed on `2026-05-21`
|
||||
|
||||
Current licensing judgment:
|
||||
|
||||
|
|
@ -1456,7 +1467,15 @@ Approved working posture:
|
|||
|
||||
- HyperTwist may use the codebase directly as the primary Python STT donor/service-layer candidate under the `MIT` code posture
|
||||
- keep it bounded behind a transcription-service seam rather than letting it shape the gameplay/runtime core
|
||||
- treat it as complementary to `ggml-org/whisper.cpp`: prefer `whisper.cpp` for harder native/runtime boundaries and `faster-whisper` for Python orchestration, batch transcription, and rapid feature delivery
|
||||
- the top-level provider/session contract remains first-party HyperTwist-owned and
|
||||
provider-neutral; this row does not own that lane
|
||||
- treat it as complementary to `ggml-org/whisper.cpp`: prefer `whisper.cpp` for harder
|
||||
native/runtime boundaries and `faster-whisper` for Python orchestration, batch transcription,
|
||||
richer VAD/clip orchestration, prompt hints, and rapid feature delivery
|
||||
- the first bounded `Phase 6R-F` implementation slice is now fixed to:
|
||||
- Python transcription-service orchestration profile
|
||||
- keep model files, downloadable payloads, and future voice-asset review separate from the
|
||||
code-license judgment
|
||||
|
||||
### `PostHog/posthog`
|
||||
|
||||
|
|
|
|||
|
|
@ -38,6 +38,13 @@ The retained owner remains:
|
|||
|
||||
- `ggml-org/whisper.cpp`
|
||||
|
||||
The top-level provider/session contract owner does not change here:
|
||||
|
||||
- first-party HyperTwist remains the provider-neutral `A1 / R1 / F1` owner for the speech
|
||||
provider-contract lane
|
||||
- this packet keeps `ggml-org/whisper.cpp` bounded to the narrower `A1 / R1 / F2`
|
||||
transcript-session donor slice only
|
||||
|
||||
The granted family remains bounded to:
|
||||
|
||||
- speech session configuration
|
||||
|
|
@ -144,9 +151,13 @@ This packet consumes the current `Phase 6R-E` implementation slice.
|
|||
|
||||
The next clean move is not another immediate `whisper.cpp` widening by default.
|
||||
|
||||
The next queue head should be:
|
||||
The later same-day provider-family backfill and control pass keeps the queue interpretation narrow:
|
||||
|
||||
- a source-backed `Phase 6R-F` `SYSTRAN/faster-whisper` preparation/control pass
|
||||
- first-party HyperTwist remains the top-level provider-neutral speech contract owner
|
||||
- `SYSTRAN/faster-whisper` may widen only as the bounded Python transcription-service
|
||||
orchestration donor slice
|
||||
- `ggml-org/whisper.cpp` does not inherit Python orchestration or broad provider ownership by
|
||||
implication
|
||||
|
||||
Keep the future sequencing guards visible:
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,218 @@
|
|||
# HyperTwist Phase 6R-F faster-whisper transcription orchestration preparation packet
|
||||
|
||||
Created on `2026-05-21`
|
||||
|
||||
## Status
|
||||
|
||||
- current preparation authority
|
||||
- bounded post-`Phase 6R-E` control slice
|
||||
- the first bounded implementation slice should land separately after this packet
|
||||
|
||||
## Purpose
|
||||
|
||||
This packet freezes the next widening order after the landed `Phase 6R-E` `ggml-org/whisper.cpp`
|
||||
speech transcript-session slice.
|
||||
|
||||
The open task was:
|
||||
|
||||
- define the first bounded source-backed widening packet for `SYSTRAN/faster-whisper` as the
|
||||
retained Python transcription-service orchestration profile above the already landed first-party
|
||||
speech session boundary
|
||||
|
||||
It is not:
|
||||
|
||||
- a full `faster-whisper` row transplant
|
||||
- a replacement for the first-party HyperTwist provider-neutral speech contract
|
||||
- a live microphone capture or device-permission packet
|
||||
- a model checkout, payload-shipping, or model-card approval packet
|
||||
- a real-time streaming or diarization packet
|
||||
- a `rhasspy/piper` or `coqui-ai/TTS` voice-output packet
|
||||
- a broad assistant-platform packet
|
||||
|
||||
## Current authority basis
|
||||
|
||||
This preparation packet stands on already-closed authority:
|
||||
|
||||
- `docs/REPO_LICENSE_TRACKING.md`
|
||||
- `docs/v6_5_deep_manual_pack/HyperTwist/ROADMAP.md`
|
||||
- `docs/ops/HYPERTWIST_PROVIDER_NEUTRALITY_AND_BYOK_DOCTRINE_2026-05-21.md`
|
||||
- `docs/arch/HYPERTWIST_PHASE6R_E_WHISPER_CPP_SPEECH_TRANSCRIPT_IMPLEMENTATION_PACKET_2026-05-21.md`
|
||||
|
||||
The key accepted routing facts are:
|
||||
|
||||
- first-party HyperTwist already owns the top-level provider-neutral speech/session contract lane
|
||||
- `ggml-org/whisper.cpp` already owns the landed offline/native transcript-session donor slice
|
||||
- `SYSTRAN/faster-whisper` is the next retained donor for Python-side transcription-service
|
||||
orchestration
|
||||
- the donor value there is strongest for:
|
||||
- batched decoding orchestration
|
||||
- VAD-aware clip orchestration
|
||||
- word-timestamp and transcript-metadata shaping
|
||||
- prompt-hint surfaces such as `prefix` and `hotwords`
|
||||
- language-detection and chunking controls
|
||||
- ownership denied there is:
|
||||
- do not let the row absorb the top-level provider contract
|
||||
- do not let the row absorb BYOK/profile custody
|
||||
- do not widen into microphone capture or permission shell ownership
|
||||
- do not widen into downloadable model or payload review
|
||||
- do not widen into TTS or broad voice-output ownership
|
||||
- do not widen into a broad assistant platform
|
||||
|
||||
## Why this was the next packet
|
||||
|
||||
The earlier retained queue heads were already landed:
|
||||
|
||||
- `Phase 6R-A` `HactarCE/Hyperspeedcube`
|
||||
- `Phase 6R-B` `kkoomen/qbr`
|
||||
- `Phase 6R-C` `vivaansinghvi07/rubix-cube-solver`
|
||||
- `Phase 6R-D` `roice3/MagicTile`
|
||||
- `Phase 6R-E` `ggml-org/whisper.cpp`
|
||||
|
||||
That advanced the live queue to:
|
||||
|
||||
- `Phase 6R-F` `SYSTRAN/faster-whisper`
|
||||
|
||||
with the later adjacent retained rows ordered behind it:
|
||||
|
||||
- `rhasspy/piper`
|
||||
- `coqui-ai/TTS`
|
||||
|
||||
## Required result
|
||||
|
||||
The source-backed control pass for this packet is now complete.
|
||||
|
||||
The first actual `6R-F` implementation packet should:
|
||||
|
||||
- preserve the first-party provider-neutral speech contract as owner
|
||||
- define one bounded retained slice from `faster-whisper`
|
||||
- keep the slice inside the accepted speech-input seam
|
||||
- widen only the first Python orchestration family that can stand on its own without dragging in
|
||||
model shipping, provider-profile custody, or voice-output scope
|
||||
- explicitly state which neighboring retained rows stay closed in that packet
|
||||
|
||||
## Source-backed retained basis
|
||||
|
||||
The queue-head decision is now source-backed rather than README-only.
|
||||
|
||||
Inspected retained donor basis:
|
||||
|
||||
- `README.md`
|
||||
- `setup.py`
|
||||
- `faster_whisper/transcribe.py`
|
||||
- `faster_whisper/vad.py`
|
||||
- `tests/test_transcribe.py`
|
||||
|
||||
Inspected first-party receiving basis:
|
||||
|
||||
- `UnrealHyperTwist/Source/UnrealHyperTwist/Public/HyperTwistRecognition/HyperTwistRecognitionTypes.h`
|
||||
- `UnrealHyperTwist/Source/UnrealHyperTwist/Public/HyperTwistRecognition/HyperTwistSpeechClient.h`
|
||||
- `UnrealHyperTwist/Source/UnrealHyperTwist/Private/HyperTwistRecognition/HyperTwistHttpSpeechClient.cpp`
|
||||
- `UnrealHyperTwist/Source/UnrealHyperTwist/Private/HyperTwistRecognition/HyperTwistMockSpeechClient.cpp`
|
||||
- `UnrealHyperTwist/Source/UnrealHyperTwist/Public/HyperTwistTraining/HyperTwistTrainingSubsystem.h`
|
||||
- `UnrealHyperTwist/Source/UnrealHyperTwist/Private/HyperTwistTraining/HyperTwistTrainingSubsystem.cpp`
|
||||
- `UnrealHyperTwist/Source/UnrealHyperTwist/Public/HyperTwistBootstrap/HyperTwistContractLibrary.h`
|
||||
- `UnrealHyperTwist/Source/UnrealHyperTwist/Private/HyperTwistBootstrap/HyperTwistContractLibrary.cpp`
|
||||
|
||||
## Narrowed 6R-F slice decision
|
||||
|
||||
The first widening slice is fixed as:
|
||||
|
||||
1. Python transcription-service orchestration profile above the existing first-party speech session
|
||||
boundary
|
||||
|
||||
not as the broader speech-input or provider platform.
|
||||
|
||||
That narrowed slice covers:
|
||||
|
||||
- first-party transcription-orchestration request policy for:
|
||||
- batching
|
||||
- chunk length / clip segmentation
|
||||
- prompt-hint routing:
|
||||
- `prefix`
|
||||
- `hotwords`
|
||||
- language-detection thresholds and segment-count policy
|
||||
- condition-on-previous-text and word-timestamp request posture
|
||||
- first-party transcript metadata shape for orchestration-aware results
|
||||
- direct-donor speech-client seam updates for:
|
||||
- mock coverage
|
||||
- HTTP-sidecar request/response normalization
|
||||
- active subsystem normalization of the bounded orchestration profile
|
||||
- focused automation coverage
|
||||
|
||||
## A/R/F clarification by slice
|
||||
|
||||
Provider contract lane:
|
||||
|
||||
- first-party HyperTwist
|
||||
- `A1 / R1 / F1`
|
||||
|
||||
Landed offline/native transcript-session slice:
|
||||
|
||||
- `ggml-org/whisper.cpp`
|
||||
- `A1 / R1 / F2`
|
||||
|
||||
Chosen next Python orchestration slice:
|
||||
|
||||
- `SYSTRAN/faster-whisper`
|
||||
- `A1 / R1 / F2`
|
||||
|
||||
These are different slices.
|
||||
|
||||
Do not let the donor that wins the orchestration slice inherit the provider-contract lane by
|
||||
accident.
|
||||
|
||||
## Deferred neighboring families
|
||||
|
||||
The first `6R-F` implementation packet must keep these capability families closed:
|
||||
|
||||
- live microphone capture shell
|
||||
- device-permission shell
|
||||
- downloadable model checkout, packaging, or payload shipping
|
||||
- worker-process management beyond the bounded request/response orchestration seam
|
||||
- real-time streaming, diarization, or broad assistant-platform scope
|
||||
- provider-profile, BYOK, or OpenAI-compatible endpoint custody
|
||||
- TTS or voice-output ownership
|
||||
|
||||
## Why this slice is first
|
||||
|
||||
- `faster-whisper` is strongest where the retained donor adds orchestration controls above the
|
||||
already landed transcript-session contract:
|
||||
- batch size
|
||||
- clip/VAD control
|
||||
- prompt hints
|
||||
- word timestamps
|
||||
- language detection
|
||||
- current first-party code already has a bounded speech session boundary, but it does not yet own a
|
||||
richer Python orchestration profile for that seam
|
||||
- keeping model, payload, provider-profile, and voice decisions outside this packet preserves
|
||||
swapability, provider neutrality, and cleaner legal review
|
||||
|
||||
## Out of scope for the first 6R-F packet
|
||||
|
||||
- microphone capture
|
||||
- downloadable model checkout or packaging
|
||||
- provider-profile and BYOK custody
|
||||
- generic provider dashboarding or usage accounting
|
||||
- TTS / voice output
|
||||
- `rhasspy/piper` or `coqui-ai/TTS`
|
||||
- broad assistant or narration platform work
|
||||
|
||||
## Acceptance criteria
|
||||
|
||||
- the packet records why the live queue advanced from landed `6R-E` to `6R-F`
|
||||
- the packet records that the first `6R-F` widening slice is the Python transcription-service
|
||||
orchestration profile only
|
||||
- the packet states exact out-of-scope families for the first `6R-F` pass
|
||||
- the packet keeps the speech-lane legal sequencing guard visible:
|
||||
- keep code-license judgments separate from model, voice, and payload-license review
|
||||
- the packet keeps the provider-neutrality guard visible:
|
||||
- first-party HyperTwist remains the top-level provider-contract owner
|
||||
|
||||
## Validation checklist
|
||||
|
||||
1. confirm the landed `Phase 6R-E` packet is the consumed prior queue head
|
||||
2. confirm the retained `faster-whisper` seam narrows cleanly to the orchestration profile
|
||||
3. confirm the next implementation packet is framed as bounded source-backed widening rather than a
|
||||
Python speech-platform transplant
|
||||
|
||||
That is the packet.
|
||||
|
|
@ -79,11 +79,11 @@ Use these as the current governing docs:
|
|||
|
||||
The next bounded move remains:
|
||||
|
||||
1. source-backed `Phase 6R-F` `SYSTRAN/faster-whisper` preparation/control
|
||||
pass
|
||||
1. bounded permissive `Phase 6R-F` `SYSTRAN/faster-whisper`
|
||||
implementation packet
|
||||
2. keep it complementary to `ggml-org/whisper.cpp`
|
||||
3. keep the first widening slice bounded to Python-side transcription-service
|
||||
orchestration
|
||||
3. keep the first widening slice bounded to the Python transcription-service
|
||||
orchestration profile
|
||||
4. keep model/payload review separate from code-license judgment
|
||||
|
||||
## Memory-specific sequencing rule
|
||||
|
|
|
|||
|
|
@ -181,7 +181,7 @@ repo.
|
|||
|---|---|---|---|
|
||||
| Speech transcript session boundary | Implemented now | landed `whisper.cpp` packet | Current bounded STT truth. |
|
||||
| Provider-backed speech session health and transcript envelopes | Implemented now | first-party speech client surfaces | Current first-party speech seam. |
|
||||
| Python transcription-service orchestration | Deep-source grounded retained | `faster-whisper` queue head | Next bounded queue move, not yet implemented. |
|
||||
| Python transcription-service orchestration profile | Deep-source grounded retained | `faster-whisper` source-backed queue head | Bounded to batching, clip/VAD control, prompt hints, language detection, and transcript metadata orchestration; not yet implemented. |
|
||||
| Voice output / narration sidecars | Deep-source grounded retained | `piper`, `coqui-ai/TTS` retained rows | Retained for later bounded work. |
|
||||
| Broad live microphone shell and downloadable payload shipping | Deep-source grounded retained | speech stack retained remainder | Explicitly deferred. |
|
||||
|
||||
|
|
|
|||
|
|
@ -62,8 +62,11 @@ Canonical discovery surfaces for roadmap interpretation:
|
|||
now landed in current code
|
||||
- `ggml-org/whisper.cpp` remains partially incorporated; live microphone capture, model/payload
|
||||
shipping, Python orchestration, and voice-output ownership stay deferred
|
||||
- the current next bounded move is a source-backed `Phase 6R-F` `SYSTRAN/faster-whisper`
|
||||
preparation/control pass
|
||||
- the later provider-family backfill keeps the top-level provider/session contract first-party and
|
||||
provider-neutral rather than donor-owned
|
||||
- the generic source-backed `Phase 6R-F` `SYSTRAN/faster-whisper` control pass is now consumed
|
||||
- the current next bounded move is a bounded permissive `Phase 6R-F` `SYSTRAN/faster-whisper`
|
||||
implementation packet for the Python transcription-service orchestration profile
|
||||
- the repo-row implementation queue is now live from that `Phase 6R-A` entry point rather than
|
||||
waiting on another first-party packet
|
||||
|
||||
|
|
@ -125,8 +128,8 @@ Current routing truth:
|
|||
- active non-live implementation-board rows: `35`
|
||||
- retained benchmark, oracle, or clean-room-later rows outside the active implementation board: `9`
|
||||
|
||||
The next bounded move is a source-backed `Phase 6R-F` `SYSTRAN/faster-whisper`
|
||||
preparation/control pass.
|
||||
The next bounded move is a bounded permissive `Phase 6R-F`
|
||||
`SYSTRAN/faster-whisper` implementation packet.
|
||||
|
||||
Queue interpretation after that packet:
|
||||
|
||||
|
|
@ -183,12 +186,24 @@ Queue interpretation after that packet:
|
|||
- Python orchestration / batching / hotword service layer
|
||||
- TTS or voice-output ownership
|
||||
- broad assistant-platform scope
|
||||
- after the `Phase 6R-E` implementation packet, keep the next queue shape explicit:
|
||||
- `SYSTRAN/faster-whisper`
|
||||
- `SYSTRAN/faster-whisper` is now source-backed and narrowed as a partially retained row rather
|
||||
than a generic future placeholder:
|
||||
- next implementation slice:
|
||||
- Python transcription-service orchestration profile
|
||||
- still deferred:
|
||||
- microphone capture shell
|
||||
- provider-profile and BYOK custody
|
||||
- downloadable model or payload shipping
|
||||
- TTS or voice-output ownership
|
||||
- broad assistant-platform scope
|
||||
- after the `Phase 6R-F` implementation packet, keep the next queue shape explicit:
|
||||
- `rhasspy/piper`
|
||||
- `coqui-ai/TTS`
|
||||
- keep the speech-input / voice sidecar legal sequencing guard visible:
|
||||
- keep code-license judgments separate from model, voice, and payload-license review
|
||||
- keep the provider-neutral speech-lane guard visible:
|
||||
- first-party HyperTwist owns the top-level provider/session contract
|
||||
- `whisper.cpp` and `faster-whisper` may win narrower donor slices without inheriting that lane
|
||||
- keep the optional `Phase 3R-G` browser comparison lane deferred unless the landed primary browser
|
||||
spatial stack exposes a real gap
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue