Files
waozi ef0de86e02
Cloudflare Pages / Build and verify the site (push) Successful in 3m29s
Integrate Breeze desktop, Android, and app updates
Use the verified dependency lock across native hosts and bundled apps. Preserve saved styles, local app reactivation, and offline packages while integrating the latest practice, Habits, Lumi, and figure preview updates.
2026-11-20 02:24:34 -03:00
..

Local female narration

The app supports en es cs de fr pt ru ja ko zh. Ming handles en es de fr ja ko zh; Russian uses Misha24-10/F5-TTS_RUSSIAN, and Czech uses Thomcles/Chatterbox-TTS-Czech. Portuguese selects FearL0rd/Chatterbox-TTS-Portuguese. Indonesian and Italian are removed from active UI, text, packs and inventories, with verified recovery copies. The inventory has 496 female cues per language: guide 32, Tai Chi 307, Sun Salutation 20, original Astral 20, Workout 29, and the six Bruce practices 88. Replacing the seven Ming languages requires 3,472 clips; all ten female inventories require 4,960, including 880 new Bruce recordings. Male recordings retain their existing provider and paths in supported languages. WHM, meditation and patterns currently have no spoken narration inventory.

ming-inventory.json is an inventory snapshot, not a generation receipt. Refresh it without changing package links:

python3 scripts/practice-narration-inventory.py \
  --media-root /mnt/storage/Projects/waozixyz/modules/practices \
  --output design/narration/ming-inventory.json

The public ming-female.json descriptor pins the official model/runtime revisions, every runtime/model/ASR artifact, the synthetic female WAV and CAMPPlus embedding, the local adapter source, and Python package versions. Female configs reference it through ming. The default generation command retains the previous provider; Ming is selected explicitly with --ming-profile. No unsupported locale is sent to Ming or silently redirected to another service.

The ignored backup/narration/ming-local.json supplies existing absolute paths. Model and ASR weights now use the shared harness store; the runtime, environment and female reference remain at their original read-only installation paths:

{
  "runtime": "/absolute/existing/ming/runtime",
  "model": "/mnt/storage/Projects/harness/models/weights/tts/ming-3.0-tts-0.5b",
  "reference": "/absolute/existing/calm-female.wav",
  "embedding": "/absolute/existing/calm-female.pt",
  "asr": "/mnt/storage/Projects/harness/models/weights/asr/whisper-small",
  "verification_models": {
    "openai/whisper-small": "/mnt/storage/Projects/harness/models/weights/asr/whisper-small",
    "openai/whisper-large-v3-turbo": "/mnt/storage/Projects/harness/models/weights/asr/whisper-large-v3-turbo"
  },
  "threads": 2
}

Use the interpreter of that installation. No command below downloads or installs models. To deliberately fingerprint an installation after a reviewed adapter or runtime change, run scripts/prepare-ming-profile.py --local backup/narration/ming-local.json --output design/narration/ming-female.json. Changing verification settings invalidates affected QA receipts. The separate verification block pins Japanese and Korean to Whisper Large v3 Turbo; the other five Ming languages retain Whisper Small. Each transcript records its actual model, revision, artifact manifest hash, language, full sample count and encoded audio hash. One recognizer is loaded at a time, including language switches. --previous-verification-profile may retain explicitly pinned prior QA identities only for locales whose model, revision and artifact manifest are unchanged. It does not accept Korean's earlier Small receipts under the Turbo route.

Removing only verification reconstructs the exact original synthesis profile. The original Ming and Whisper adapter files remain unchanged, so cached PCM, request hashes and original synthesis provenance survive the verifier update. Use --cached-only with the generation command to recheck existing pilot PCM; it refuses missing or changed cached audio before any new synthesis. Fresh ASR, speaker, duration, direction, number and ending validation still runs.

Attempt diagnostics are saved under ignored backup/narration/diagnostics/<locale>/<key>/<run>.json. Each bounded record retains the actual attempt's duration, speaker threshold, ASR identity, raw text, scores and rejection reasons. A final failure distinguishes the best transcript from the last attempt. These records never certify a recording by themselves.

The Sun Salutation comparison permits 向后脉/向後脈 and 向后卖/向後賣 only when its source contains 向后迈, plus that cue's traditional script equivalents. The raw ASR text stays intact, and pulse words elsewhere are unchanged. The MOE dictionary lists the shared pronunciation for 邁 and 脈. Its 賣 entry also records that pronunciation. Japanese and Korean have no new pronunciation aliases: Turbo correctly recognizes the saved Japanese leg instruction and Korean breathing cue that Small misrecognized.

Very short Sun Salutation cues have an explicit context path for pt cs ru es fr zh. short-context-provider-v4.json binds each exact short phrase to a longer existing practice instruction, the helper and verifier sources, and the blind alignment model. Ordinary cues keep their original synthesis requests and receipt identity. Use both options for context generation or preflight:

--context-profile design/narration/short-context-provider-v4.json \
--context-local-profile backup/narration/context-local.json

The current helper/verifier use explicit successor descriptors: Sun Salutation short-context-provider-v4.json, Czech workout workout-czech-context-provider-v4.json, Tai Chi tai-chi-short-context-provider-v4.json, Czech counts workout-czech-count-semantics-v3.json, and Portuguese workout workout-portuguese-context-v3.json. Historical descriptors retain their original source seals. Existing accepted proofs need fresh actual alignment before they can migrate to a changed helper; the helper-successor staging command plans exactly fifteen parent alignments and two Chinese crop transcriptions without running inference or writing live receipts.

Portuguese counts opt in separately through workout-portuguese-count-semantics-v1.json. Only the exact registered workout cues and parents equate standalone ASCII 1, 2, 3 with um, dois, três. Raw ASR remains unchanged. Signed, decimal, ordinal, combined and non-ASCII number forms, missing or extra words and the start cue receive no aliases. The same sealed comparison reaches final assessment and receipt replay; existing speaker, prefix, duration, ending and complete decoding gates remain.

The parent instruction must pass complete blind transcription and the original provider's speaker threshold. Actual word timestamps select one contiguous prefix; the encoded crop must contain exactly the short phrase, stay within the existing speech budget and remain below full scale. Its receipt retains the actual short speaker score separately from the parent's score. The original female reference, model, thresholds and raw transcript remain intact.

The two configured Chinese short cues use separately pinned Turbo verification for the final crop. Their receipts record the actual model, revision, artifacts, audio hash and full sample coverage under the context recipe. This applies only after the complete parent/crop proof validates; ordinary Ming routes and its public synthesis profile remain unchanged. Chinese exhale uses the existing plank instruction as its parent, preserving the spoken short cue 呼气。.

Accepted parent proof WAVs are copied as binary files into sunsalutation/voice-context/<locale>/<sha256>.wav in the practices package, outside the playback and voice-pack globs. The proof binds the original synthesis identity, parent hash, word alignment, crop range and actual encoded audio. --check verifies the public proof and receipt without private weights or caches; supply only --context-profile alongside the normal provider selection. Fixture tests verify the protocol; each configured real cue still needs its own successful generation and current-receipt validation.

An explicit context candidate may use the normal four attempts when its parent fails speech quality or its exact crop fails quality checks. Each rejection retains the actual parent/crop reason, scores and identity, including failures before a short clip exists. Changed model/reference/source hashes, invalid configuration and system errors remain fatal. Ordinary provider retry behavior is unchanged.

Before generation, validate all actual file hashes without loading a model:

PYTHON scripts/generate-lumi-guide-narration.py \
  --practice astral_projection_bruce \
  --ming-profile backup/narration/ming-local.json --preflight

Replace PYTHON with the existing runtime interpreter. To generate the first English and German pilot cues, remove --preflight and add --locale en --locale de --key astral_text_foundation_01 --jobs 1. Only the media task runs generation; coordinate the shared runtime's resources and package-link ownership first. After actual foundation pilots in all seven languages pass, omit --key and --locale for all seven supported locales. Prioritize Bruce and original Astral, then guide, taichi, sunsalutation, and workout. The same command with --check verifies selected receipts offline. Male inventories reject --ming-profile.

Russian, Czech and Portuguese select a provider explicitly, using the Python interpreter recorded in that provider's ignored private JSON. Their incompatible synthesis environments remain separate; shared QA runs in its own pinned environment. Use f5-russian with backup/narration/local-models/ru-local.json, or chatterbox-czech with backup/narration/local-models/cs-local.json, or chatterbox-portuguese with backup/narration/local-models/pt-local.json:

PROVIDER_PYTHON scripts/generate-lumi-guide-narration.py \
  --practice astral_projection_bruce --local-provider f5-russian \
  --local-voice-profile backup/narration/local-models/ru-local.json --preflight

Replace PROVIDER_PYTHON with the JSON's python path. Generation rejects a different interpreter before loading models. To generate a pilot, replace --preflight with --key astral_text_foundation_01 --jobs 1. --check validates receipts without an installed runtime or private model files. A provider never receives another provider's locale or falls back to a remote service. Existing clips from other providers/locales remain in mixed receipts. Actual provider pilots must pass before bulk generation; descriptor status alone is insufficient.

One process caches one Ming model and one Whisper model. Each synthesis uses explicit language_id, FREEFORM calm delivery and the same nonzero female speaker embedding. Text hashes bind the complete request, language, seed, model, reference, runtime, adapter and output encoding. Atomic per-cue receipts resume completed work. Ignored backup/narration/cache/ preserves exact generated PCM requests across interrupted QA, avoiding another synthesis of the same request. The output remains mono Ogg Opus at the existing app paths.

Ming's dense runtime shares a 512-token cache between the input and generated audio. The original German foundation request left only 23.28 seconds of audio budget. narration_segments.py now preserves exact source spans and splits long cues at sentences, clauses and safe whitespace, targeting 24 semantic units. Short cues remain one request; words, decimal numbers and Unicode whitespace are preserved. Every span uses the same original female embedding and explicit language, with a deterministic seed. Float waveform caches resume completed spans. Assembly concatenates their audio without overlap or added silence. The profile and request identity bind the segmenter, plan and assembly settings. Every actual prompt must leave at least 16 seconds of decode budget before synthesis. Whole-cue QA remains required after assembly. Finite raw waveforms are saved as float WAV before peak rejection; empty and nonfinite results are rejected with numeric diagnostics. No automatic clipping or gain is applied.

Every accepted clip must be finite and unclipped, fit its speech budget, retain the reserved Bruce quiet interval, match its complete transcript, and meet the recorded CAMPPlus speaker-similarity bound. Sun Salutation also fits its shortest six-second default breath with a 0.2-second margin. Whisper receives all samples without truncation and uses its sequential timestamp-based decoder, including cues longer than 30 seconds. QA supplies no expected-text prompt. Every local cue also requires matching final 64- and 24-character transcript endings, using the same locale aliases as the whole-cue comparison. This rejects the initial 23-second German pilot's missing final clause despite its 0.92958 overall score; the complete English pilot remains accepted. Receipts bind QA to the actual encoded audio hash and full decoded duration. Speaker similarity measures consistency; it does not establish how a listener perceives the voice.

python3 tests/ming_narration_test.py tests hashes, request identity, rejected locales, real Opus fixture encoding, quiet intervals and full long-audio routing. Fixtures are separate from production media and do not certify real Ming speech. python3 tests/local_narration_pipeline_test.py verifies separate provider routing, wrong-interpreter rejection and full receipt identity. These local tests use isolated fixtures. tests/ming_segment_pipeline_test.py, run under the existing Ming interpreter, verifies real float WAV, PCM assembly and cache resume without loading a model. Real seven-language Ming pilots and all selected Russian/Czech/Portuguese pilots, all recordings and final native playback remain required before the feature can be committed or described as complete. No partial pack may advertise missing languages as complete.

Official references: Ming controls and Whisper long-form decoding.