# Voiceclone — Backlog / To-do Things noted to come back to later. ## ⏰ REMINDERS — come back to these (parked 2026-07-17) - [ ] **Ear-test the redesigned Broadcast polish.** Rebuild "Engineered" (project 68) on **Broadcast level** — should sound like the raw voice at proper loudness, zero underwater processing. If clean → it's the safe default for launch. - [ ] **Ear-test De-noise on the ElevenLabs voice with audible take-joins** (e.g. "DK eleven 812pm"). Does it hide the seams without dulling? Report back so the strength gets tuned by verdict, not guesswork. - [ ] **Promo producer parked** — beds/SFX not good enough for clients yet (see Promo v3 item: real sample packs are likely the answer). - [ ] **Launch target moved (was week of 2026-07-20, no new date).** Before launch: pick polish default, Stripe payments decision. ~~Hide lab links from client nav~~ → DONE 2026-07-23: internal links (⚡ Clone Voice shortcut + 🧪/🎚/🗣/📻 labs) now show only for accounts in config `app.internal_emails` (currently test@voiceclone.dev + dave@davekirwan.com.au) — clients never see them in any environment. ## Quality / polish - [ ] **Emotion & whisper are too subtle** (raised 2026-07-08). ElevenLabs v3 audio tags (`[whispers]`, `[excited]`, `[sad]`, …) work but the effect is weak. Ideas to strengthen later: - Tune ElevenLabs `voice_settings` per generation (lower `stability`, raise `style`) — more expressive but less consistent. - Try stronger/edgier tag wording and placement (e.g. `[whispering quietly]`). - Compare voices — some ElevenLabs voices emote far more than others. - Revisit once we're on a higher tier / newer model. ## Delivery / naturalness - [x] **Audio tags + punctuation surfaced to the client** (2026-07-10). Added the "Shape the delivery" cheatsheet on the editor + a one-click **✨ Perform it…** palette (whisper/softly/slowly/sigh/warmly/excited/laugh/trail-off) that writes the tag into the line via edit-word. Relabelled the old "Pause after" → "Trail off …" (it already injects an ellipsis into the ElevenLabs text). Everything typed (`…`, `[tags]`, CAPS, `—`) passes through verbatim (verified). - [x] **AI Auto-Director** (2026-07-10) — BUILT. `public/auto-direct.php` + `src/anthropic.php` send the whole script + preset to Claude (config `anthropic.model`, default claude-sonnet-5) and return a directed version of each line (tags + `…` + emphasis added, words untouched — verified by a word-equality guard that rejects any line where the model changed the words). Editor button "✨ Auto-Direct my script" → review modal → apply via edit-word. Free (no credits). **Needs `config.anthropic.api_key` set to switch on.** TODO later: meter/limit usage (real Claude cost); optional per-use credit charge; let the client tweak intensity (subtle ↔ theatrical). - [x] **Director intensity + wider tags** (2026-07-10). Director has a Subtle/Balanced/Theatrical chooser (setup step in the modal, each clearly explained) passed as `intensity` to auto-direct. Cheatsheet + director vocabulary expanded with breaths/laughs/etc. Working live. - [x] **Removed the "what the voice receives" preview** (2026-07-10) — it named ElevenLabs to clients; user doesn't want the provider exposed. Feature fully stripped (toggle, per-line preview, `body` returns). See no-expose-provider rule. - [x] **Inline line editing (document-style)** (2026-07-10). Each line has an "✏️ Edit" button → swaps the rendered `

` for a textarea of the raw text so the client can type freely (add `…`, CAPS, `[tags]`, delete director tags they don't want). Toolbar: "Add …", "CAPS" (uppercase selection), Save/Cancel, Ctrl+Enter=save, Esc=cancel. Saves via edit-word (full-line replace); edit-word + nuance now return `raw` so the edit-box stays in sync. `.seg__text` carries `data-raw`. Root cause the user hit: cheatsheet said "type into a line" but lines were never editable. Also click-to-edit: a plain click on the text opens the editor with the caret at the click point (caretRangeFromPoint); drag-selecting still opens the highlight menu instead. ✏️ Edit button kept for discoverability. - [x] **Director mood guardrail + preserve client markup** (2026-07-10). (Correction: the `[laughs]`-in-meditation was the USER testing manual tags, not a director mistake.) Prompt now (a) preserves any tags/…/CAPS the client typed themselves — never removes or moves them, and (b) forbids the DIRECTOR from adding laughter/ playful sounds to calm/meditative/sleep/serious/sad/corporate/luxury content. Verified live: user tags kept in place, additions stay mood-appropriate. - [ ] **Go deeper on delivery** (later): per-line "ending" quick-set; curated tag groups per use-case; director intensity could become a saved per-project default. - [x] **Lead-in/out fades were too heavy** (2026-07-10) — big volume ramps sounded like swelling/ducking. Cut to subtle click-guards; naturalness now comes from punctuation (`…`) not volume. See `src/presets.php` fade_in/fade_out. ## Audio - [x] **Rebuild ignored edits + build history** (2026-07-10). Root cause: rebuild re-stitched the OLD takes — editing a line never regenerated its audio, so rebuilds sounded identical. Now: builds are numbered and all kept (`project_builds` table, files `project-{id}-build-{n}.mp3`, UI list newest-first with LATEST pill + per-build download `download.php?build=n`). Stale detection: `text_marked IS NULL` + selected take = line changed since generated (edit-word already nulls it; nuance.php now nulls it too). export.php returns `needs_generation {missing, changed, lines[]}`; JS flow is server-driven (computeMissing/missingSegs removed); confirm box offers "Build for N cr" or "Build with old audio" (`force_stale`). Verified live: BUILD 1→3 on proj 23, stale flag message, force-stale path. - [x] **End-of-clip cut-off** — fixed by the custom player (plays to true `ended`). - [x] **Clicks at joins + "digital" quality** (raised 2026-07-10) — fixed: switched takes to **uncompressed pcm_24000** (v3 native, no MP3 artifacts at source), stitch now does a single 192k MP3 encode with 12ms edge-fades per line. Old MP3 takes need re-generating to benefit. - [x] **Pro tier + pcm_44100** (2026-07-13) — user upgraded to ElevenLabs Pro; `output_format` bumped to `pcm_44100` and verified live (WAV header = 44100 Hz). Takes are now full-bandwidth uncompressed end-to-end until the final 192k MP3. Old takes are still 24k — regenerate lines to upgrade them (stitch mixes both fine). - [ ] Offer a **lossless WAV download** option for final builds (chain is lossless up to the MP3 encode now — stitch could emit a WAV alongside). - [ ] **Per-line ending inflection control** (raised 2026-07-10). v3 doesn't accept surrounding-line context (`previous_text`/`next_text` → HTTP 400), so delivery varies line to line. Give the client an easy per-line "Line ending" dropdown (Statement `.` / Trail off `…` / Question `?` / Continue `,`) that adjusts the trailing punctuation sent to ElevenLabs — the main lever for rising vs falling inflection on v3. Pairs with the new preset lead-in/out fades. ## Phase 3 — remaining - [x] **Stitch-to-download** — DONE. `public/export.php` + `src/stitch.php` stitch each segment's selected take with its pause (ffmpeg at C:\Users\davte\ffmpeg), `public/download.php` serves it named after the project. Falls back to plain concat if ffmpeg is unavailable. - [ ] Word-level pronunciation is single-word only; add multi-word phrase support. - [x] AI Auto-Director — DONE (see Delivery/naturalness section). Working live. ## Client voice cloning (built 2026-07-10) - [x] **Client voice cloning — BUILT.** Client reads a range-covering script, records (MediaRecorder) or uploads, consents, and we Instant-Voice-Clone it. Files: `voice-new.php` (page + read-aloud script), `assets/js/clone.js` (record/upload), `clone-create.php` (convert→`/v1/voices/add`→store), `clone-delete.php`, `src/clone.php` (ffmpeg convert + guards), `elevenlabs_add_voice`/`elevenlabs_delete_voice`. Schema: `voices.owner_client_id` + `voice_type='client_clone'` (migration applied + schema.sql). voices.php shows "Your voices" (cap 3) + delete; project-new ownership-scoped. Free to clone; generation uses CLONE_PRICE_MULTIPLIER (1.5×). No provider name shown to clients. - [ ] **Cloning needs a fully-permissioned key + cloning-capable plan.** Current ElevenLabs key looks TTS-scoped (401 on `/v1/user/subscription`) — the live clone call (`/v1/voices/add`) needs a key with **Voices write** permission AND a plan with Instant Voice Cloning (Starter+/Pro). Untested end-to-end until then. - [x] **Project archive → delete** (2026-07-10, needed to free cloned-voice slots). Two-step safety: dashboard rows have an "Archive" button (reversible, keeps all audio, no confirm) → project moves to an "Archived" section below with "Restore" and "✕ Delete forever" (confirm). `project-archive.php` archives/restores (restore derives status: completed/in_progress/draft); `project-delete.php` REFUSES non-archived projects (?err=notarchived), then cascades segments/takes/builds, clears moderation_flags (no FK), audit-logs, and unlinks all take/build/final audio files from disk. Dashboard lists ALL projects (12-row cap removed). Clone-delete "in use" flash links to the dashboard. NOTE: archived projects still reference their voice — a clone frees up only after its projects are deleted forever. Verified end-to-end on a throwaway project. - [x] **Studio-quality recording input** (2026-07-10, user's Sennheiser sounded awful). Root cause: browser defaults (echoCancellation/noiseSuppression/ autoGainControl) mangle good mics + MediaRecorder compressed to webm/opus. Now: clone.js captures RAW via WebAudio (all processing off), encodes lossless 16-bit WAV in-browser; mic picker (enumerateDevices), INPUT GAIN slider (GainNode, post-gain metered), live INPUT LEVEL meter (green/yellow/red + peak-hold line), clearly labelled. Server converts to mono 44.1k WAV (not mp3) and sends lossless to the cloner; `silenceremove` trims leading silence keeping 0.5s (verified: 3s silence+2s tone → 2.5s). Custom aplayer replaces the browser audio element (shows length); Record button no-wrap. - [x] **Accent drift fix** (2026-07-13, Australian clone read American then drifted). Three levers: (1) clone creation now sends `labels` {accent, gender} + a `description` to the provider — anchors identity so the model stops guessing per line; clearly-labelled Accent (default Australian) + gender selects on voice-new (stored in voices.accent/gender). (2) generation with `client_clone` voices sends `stability: 1.0` (Robust — closest to reference; v3 accepts only 0/0.5/1; trade-off: audio tags slightly subtler on clones). (3) guidance now says 60–90s ideal, read the script twice. Existing unlabelled clones must be deleted + re-cloned to gain the anchor. - [x] **Professional Voice Cloning** (2026-07-13, Pro tier live) — BUILT. Wizard at `voice-pro.php` (+ `assets/js/pvc.js`, endpoint `public/pvc.php`, provider fns in elevenlabs.php): create (accent/gender labels + consent) → upload 30min–3h of samples (200MB/file; php.ini now 200M/220M — NEEDS SITE RESTART) → speaker verification (provider captcha image read aloud, raw-WAV recorder) → train → poll `fine_tuning.state` until `fine_tuned` (status action flips is_active=1). `voices.provider_state`: pvc_draft→pvc_verified→pvc_training→NULL. Voices page shows in-progress cards ("Setup incomplete"/"Training…") + ⭐ Professional clone CTA; one PVC in progress per client. Smoke-tested live: provider create + card + delete round-trip OK on the Pro key. Untested with real audio: sample upload at scale, captcha verify, training completion — user will run the first real one. NOTE: provider may not have PVC fully optimised for the v3 model yet — first real generation will tell; Robust stability already applies (client_clone). - [x] **Browser-side 44.1k downsample before upload** (2026-07-13). User's interface records at 96/192kHz → huge WAVs kept tripping upload limits (his 1:24 take failed INI_SIZE because his site window predates the 200M ini bump). Both recorders (clone.js + pvc.js createRecorder) now resample captured PCM to 44.1kHz via OfflineAudioContext before WAV encode — ~4× smaller, zero audible loss (44.1k is the delivery rate anyway), immune to server limits. Verified in engine: 192k→44.1k exact frame count, RMS + 440Hz preserved. Mic hint now says "saved as lossless 44.1 kHz WAV". - [x] **Instant-clone script doubled + 7th style** (2026-07-13, user spec). Each read-aloud line ~2× longer (~27–35 words each, ~226 words ≈ 90s total) and a new "natural storytelling" style added for connected-speech coverage. Recording caps raised to match: clone.js MAX_SEC 90→180, CLONE_MAX_SEC 120→180. Hint now says one read ≈ 90s = the sweet spot; fluffed lines: pause and re-read the line. - [x] **PVC audio range: 1 min – 1 hr** (2026-07-13, user spec). Step 2 gates "Continue to verification" until ≥60s total; hard 1-hour ceiling disables record/upload with a clear message; progress bar full point stays at the 30-min "ideal". Resume now honest: new `pvc.php?action=samples` lists already-uploaded samples from the provider (file_name + duration) and repopulates the total on reload. Verified live (draft resumes at step 2, "0 sec", gate text shown, cleanup OK). - [x] **20MB cap fix + PVC browser recording** (2026-07-13). The "too large (max 20 MB)" error was the app-level CLONE_MAX_BYTES cap — sized for 48kHz WAV, but pro interfaces record at 96/192kHz (90s @ 192k ≈ 35MB). Cap raised to 200MB everywhere (clone.php const, clone-create messages, clone.js pre-check). PVC step 2 now records in-browser too: "● Record a take" (raw WAV, 10-min takes, auto-uploads as browser-take-N.wav) alongside file upload; recorder refactored into createRecorder() factory shared with the verification step. - [x] **Clone-faithful rendering + PVC slot recovery** (2026-07-14). (a) v3 is unfaithful to instant clones (user's Australian voice rendered British despite a verified-good 62s sample + correct labels) → cloned voices now synthesize on `eleven_multilingual_v2` (`clone_faithful` flag in generation → tts.php; config `elevenlabs.clone_model_id`), [tags] stripped on that path, fidelity settings (stability .65 / similarity .9 / speaker boost). (b) Provider REFUSES API deletion of professional voices (403) → deleting the local card orphaned "DK pro" which held the 1/1 PVC slot → re-adopted as local voice 158 (pvc_draft); wizard verified resuming at step 2 with the 62s sample. elevenlabs_delete_voice now logs non-2xx loudly. - [ ] PVC deletes: surface "the provider kept the professional voice (support ticket needed)" to the client instead of a clean 'deleted' flash; offer re-adopt automatically if an orphaned PVC is detected at create time. - [ ] Cloning polish later: charge credits per clone (currently free); show "in use by N projects" before delete; "voice match strength" control (Robust vs Expressive) per project; PVC sample management (list/delete at provider); manual-verification fallback if captcha keeps failing. - [x] **EL library voice → roster + drift fix** (2026-07-15). Added "Australian Male Voice - Dave" (BxwgGCmcalNbSzNt0b2l) via /v1/voices/add/{owner}/{id}; voices row 166. Preview was AMERICAN: sample.php synthesizes via v3 (drift!) — replaced cached sample with the library's official recorded preview_url and reclassified the voice as actor_clone; generation.php clone_faithful now covers actor_clone too (v2 rendering). voices.php roster shows stock + actor_clone (owner NULL). - [x] **Third delivery dialect: "plain" for faithful clones** (2026-07-15, user hit it: [tags] on Dave = normal read). EL client/actor clones render on the accent-locked engine which strips [tags], so projects on those voices now get: punctuation-only cheatsheet ("directed with pacing and punctuation"), no ✨ Perform-it menu, emotion dropdowns disabled (inert there), and the Auto-Director prompt forbids all bracketed/parenthetical cues (pacing only). $dialect = fish | eleven | plain, computed in project.php + auto-direct.php. Verified: Dave project = plain UI; Fish project unchanged. - [ ] Library-add polish: when adding future EL library voices, auto-download the official preview_url at add time (never synthesize previews for clones). - [x] **Voice isolation lab** (2026-07-15). Internal page /isolate-test.php (header "🎚 Isolate"): upload a mix (voice + music/effects, ≤200MB/≤30min) → ElevenLabs /v1/audio-isolation (elevenlabs_isolate in src/elevenlabs.php, ~1000 chars quota per audio-minute) → A/B players (original vs voice-only) + download. Endpoint /isolate-api.php keeps originals + results under /media/isolate/. VERIFIED with real API: loud synth pad over speech — music-only tail −24.7dB → −49dB (≈24dB kill) with speech intact. If it becomes a client feature: charge site credits per minute + queue long files. - [x] **Speaker separation lab** (2026-07-15). Internal /separate-test.php (header "🗣 Speakers"): upload interview/podcast → Scribe diarized STT (elevenlabs_transcribe_diarized, model scribe_v1, diarize=true, word-level speaker_ids) → per-speaker segments (0.6s join-gap, 60ms pad, cap 6 speakers / 400 cuts) → ffmpeg atrim+concat one MP3 track per speaker + transcript snippet so you can tell who's who. Timeline-based (not un-mixing — overlap goes to the dominant speaker). Test conversation at /media/separate/test-conv.mp3 (Dave↔Lily alternating). ⚠ BLOCKED on key scope: the EL key needs the **speech_to_text permission** enabled in their dashboard (isolation + TTS + voices already work) — error surfaces cleanly until then. - [x] **Radio promo producer lab** (2026-07-15, user request: dry VO over 90s zaps/zings). /promo-test.php (header "📻 Promo"): source = any project build (dropdown) OR uploaded dry VO; 3 styles (90s Power Promo / FM Sweeper / Chart Show); src/promo.php synthesizes ALL SFX with ffmpeg lavfi (aevalsrc chirp zaps/zings, noise boom/whoosh, sine-chord bed with tremolo pump) + VO chain (compress, presence EQ, slapback aecho) + bed DUCKED under voice via sidechaincompress (key must be apad-ed or it truncates the bed!) + alimiter + fade. Local + free (no provider quota). Verified end-to-end: build → 33s promo in 2.1s, all 3 styles measure correct (hot intro/ducked bed/fade to −60dB). - [x] **Promo producer: 8 formats + echo control** (2026-07-15, user: 30s read had way too much echo). Styles now data-driven in src/promo.php: CHR/Top 40, Soft AC, 90s Power Promo, FM Sweeper, Chart Show, Rock Station, Urban/Hip-Hop, News/Talk — each with its own bed chord/pump + intro/outro SFX. "Echo on the voice" select (No/Light/Medium/Big, default Light, plain-English descriptions + hint that long reads want light/none). Echo gains fixed constant (in/out 0.9) so the VOICE level no longer changes with the echo setting — only the repeats do (verified by null-testing renders; ladder ordering correct, level differences huge between light and big). All 8×4 combinations render. - [x] **ROOT CAUSE of "echo the whole way through"** (2026-07-15). The promo filtergraph referenced `[outro]` in the final amix but NEVER DEFINED it ($outro built, never concatenated into $fc). ffmpeg does **not error** on a dangling label — it silently mis-wires the graph, so the RAW un-ducked, un-volumed bed leaked into the mix. Symptoms: "Music none" and "Music loud" measured identical (440Hz in a pause: −27.3 vs −27.1dB); a silent pause read −15.5dB vs the source's −77.3dB. Fix = add `$outro . ';'` to $fc. After: pause = −74.6dB (≈ source), bed ladder none/quiet/balanced/loud = −91/−48/−44/−41dB. LESSON: every label referenced in the final amix must be defined in the graph — verify by dumping $fc, not by trusting exit code 0. - [x] **Selective echo (send/return)** (2026-07-15, user: echo should be sparing, on key words/lines only). Echo is now discrete taps [ms,gain] mixed ON TOP of the untouched dry voice — dry level/timing never change. `promo_echo_places()`: Ending only (default) / Opener + ending / Whole voiceover. Verified by region nulls: ending-only → mid-read −42dB (dry) vs tagline −0.9dB (echo); whole-VO mid-read −7.7dB. Earlier concat approach caused timing drift (opener mid-read nulled 0.0dB) — replaced. - [x] **"No echo still echoes" investigated (round 1)** (2026-07-15). Null-tests PROVED echo=off is honored (page render vs CLI off = −91dB in voice region, bit-identical; noise SFX are unseeded so whole-file nulls mislead). The perceived "echo the whole way through" = the BED swelling back after every word (sidechain release 350ms → voice-shaped swells). Fixed: release 900ms + attack 6 (bed stays down through phrases) and a new "Music volume" select (quiet .5 / balanced .8 / loud 1.15 post-duck gain; ladder verified by region null-tests). UI: third select on /promo-test.php, JS sends `bed`. - [x] **Promo engine v2: real drums + arrangement + 3 takes** (2026-07-15, user: "too basic, no pumping bed, needs punch, be SUPER CREATIVE"). Beds are now RHYTHMIC — synthesized kick/snare/hat/bass via `mod()`+`exp()` envelopes in aevalsrc (verified: 90s kick pulses −13dB on-beat → −27dB off-beat every 0.5s). Per-genre patterns: CHR 128bpm 4-on-floor+claps+stabs, power90s 120bpm kick/snare/16th-hats, rock 100bpm heavy, urban 808+trap hats, chart fast, news ticker, softac/sweeper padded (no drums). SFX vocabulary per style (zap/zing/riser/impact/crash/whoosh/chime/sting/stab) placed THROUGHOUT the read by `_promo_arrange()` — opener hit, 2-3 mid-read punctuations (risers pre-empt the moment they build into), impact under the last word + ring-out. **Each generate renders 3 VARIANTS** with different arrangements (verified distinct); API returns `mixes[]`, UI lists TAKE 1/2/3 each with player + download. All 8 styles × 3 variants render; live E2E ~6.8s for 3 takes. - [ ] Promo producer v3: user-uploaded bed/SFX packs (real broadcast imaging sounds), client-facing version with credit pricing ("Produce as radio promo" button on project builds), maybe per-take "regenerate this one". ## 🎬 Casting Call (launch flagship, built 2026-07-17) - [x] **The 60-Second Casting Call.** /casting.php (nav "🎬 Casting Call"): paste a script → first 250 chars auditioned across 6 diverse roster voices (actor clones first, then gender-alternating stock). Protection stack per user spec: FREE but 1/day per account (slot claimed pre-render, returned on total failure); audible watermark looped over every audition (15s unit: chime@0.5 · chime@5.5 · spoken "Audition." @10.5 — asset public/assets/audio/wm-loop.mp3, mixed at 0.4); lo-fi render 22.05kHz mono 48kbps; files in private/auditions/ OUTSIDE web root, streamed via audition-audio.php (session + ownership check, Range/206, no download disposition, anonymous = 401). Star-to-shortlist board (starred float up); "Use this voice →" → project-new.php?cast=ID prefills the FULL script. Tables: casting_calls (script + results JSON; reload restores today's board). VERIFIED on the user's own real casting call (6 voices, watermark measured in-file, limit enforced, anon blocked). - [ ] **Casting Call v2 — PRIORITY (user, 2026-07-17): extra casting calls for credits.** After the free daily call, let clients run more by spending credits — user wants this designed/built "in the next few days". - [ ] Casting Call v2 (rest): shareable read-only board link for the client's client (agency approval flow); audition credits roll into the project as a discount; pick-your-panel (choose which 6 voices); per-voice "regenerate". - [x] Casting Call filters + watermark polish (2026-07-17): gender (mixed panel / female-only / male-only) + language selects (roster-driven, lang_info labels; filters checked BEFORE the daily slot is claimed); watermark loop rebuilt — clean first ~2.5s, then strictly ONE mark every 5s (chime@2.5 · chime@7.5 · spoken@12.5, verified −91dB between marks); unmissable amber disclosure box on the form + note on the board. ## Broadcast quality (raised 2026-07-15: output not broadcast-clean; audible take joins) - [x] **Broadcast polish (mastering stage on the build)** (2026-07-15). Root cause of "you can hear the sentences join": each line is a separate AI take with its own faint baked-in room tone; stitched, that ambience blinks on/off at each boundary. Fix = master the WHOLE assembled program in one pass (not per-take) in src/stitch.php after concat: highpass → afftdn denoise (kills varying room tone → uniform floor → joins vanish) → deesser → gentle acompressor → loudnorm I=-16 LUFS (EBU R128) → TP limit. Strength select Off/Light/Standard/Strong on the build bar (default Standard); `polish` param through export.php. Safe on clean voices (denoise only acts on noise present; loudnorm+comp is universal). VERIFIED: 4 levels render 0.7–1.4s in the real build path; loudness off=-13.9 vs light/standard/strong = -16.2/-16.4/-16.3 LUFS; speech intact. Denoise sweet-spot is an EAR call → that's what the control is for. - [x] **Broadcast polish REDESIGN — was over-processed** (2026-07-17). User: Standard "sounds TERRIBLE" on a clean Fish clone. Root cause: I stacked aggressive afftdn (nr=15) + heavy acompressor + deesser on already-clean AI voices → "underwater"/ dull (old chain dropped 7kHz HF to -38dB via the deesser). LESSON: AI/clone voices are already clean+full-band — they need LOUDNESS, not scrubbing. New levels: Off / **Broadcast level (default)** = loudnorm-only, transparent, -16 LUFS / Level+polish = + gentle glue comp (ratio 2, soft knee) / De-noise = opt-in gentle afftdn nr=9 for genuinely noisy voices only. Verified: HF preserved (level=-33.7 vs raw -35.0, i.e. not dulled), renders 0.8-1.5s. Default key changed to 'level' in export.php + project.php + editor.js. User to re-test by ear. - [ ] Broadcast polish v2: per-take loudness MATCH before stitch (kills line-to-line level steps at joins — the other half); optional arnndn/RNNoise model for higher-quality denoise (needs a .rnnn model file); expose LUFS target (-16 stream vs -23 broadcast). ## Provider strategy (raised 2026-07-14: need 100+ clone slots, all languages) - [x] **Fish Audio test lab** — internal bake-off page at `/fish-test.php` (not in any nav): record (studio mode → 44.1k WAV) or upload a sample → instant Fish clone (`POST /model`, train_mode fast, private) → generate takes (`POST /v1/tts`, engine header s1/s2-pro/s2.1-pro, wav 44.1k) stacked with latency + size for side-by-side listening; refresh/delete voices. `src/fishaudio.php` + `public/fish-api.php` + `assets/js/fish-test.js`; key in config `fish.api_key` (blank → setup card). Slot economics: ElevenLabs Pro/Scale/Business = 1/3/10 PVCs (100 ⇒ Enterprise); Resemble Flex ≈ $5/mo per pro clone; Fish = unlimited instant slots on $75/mo Pro, 15 pro clones on Max. ElevenLabs leads languages (70+ on v3) — "all languages" cuts against leaving entirely. - [x] **Fish won the listening test → integrated into the main site** (2026-07-14). `FishTtsProvider` in tts.php (emotions → S1 markers: whispering/excited/sad/ confident/delighted; [tags] stripped; prosody.speed from preset; engine via config `fish.engine` = s1). Generation routes PER VOICE (`v.provider` — fish → Fish, else config default; stock roster stays ElevenLabs). NEW instant clones go to Fish via config `tts.clone_provider = 'fish'` (clone-create branches; description carries accent/gender). clone-delete branches per provider. Instant-clone cap now excludes the in-progress PVC. Existing Fish model "Dave AU test" adopted as voice 160, usable immediately. Live-verified: whisper marker + 0.8 speed → 44.1k WAV. - [x] **Delivery dialect per provider** (2026-07-14, user spotted the cheatsheet still taught [square tags] on a Fish project). Everything now switches on the project voice's provider: cheatsheet (Fish version teaches (whispering)/(soft tone)/(sighing)/(laughing)/(serious)… from the official S1 marker list, with "use (round brackets) with this voice"); ✨ Perform-it palette inserts (markers) for Fish ([tags] for EL; fish swaps slow-down→serious since S1 has no pace marker); Auto-Director prompt has a Fish variant (parenthetical vocabulary, "never square brackets"); _director_words strips (markers) too (verified); markDirection highlights Fish markers. FishTtsProvider emotion map corrected to official markers: serious→(serious), warm_smile→(comforting). Verified live: project 50 (Fish) shows round-bracket UI; project 23 (EL) unchanged. window.VOICE_PROVIDER injected by project.php. - [x] **Fish upload TLS failures fixed** (2026-07-15). "bad record mac" on clone uploads: this network kills TLS uploads between 6–9MB (https-inspection AV; HTTP/1.1 + TLS1.2 caps + retries all in place and insufficient — measured with probes). Fix = `fish_prepare_upload()` in src/fishaudio.php: every Fish upload >5.8MB or >265s is transcoded to mono MP3 (bitrate 96–192k picked to fit budget) and trimmed to 265s — Fish hard-rejects references over 270s ("Total duration of audios is too long... less than 270 seconds"). Verified: a 40MB WAV that failed 100% now clones in ~60s end-to-end. NB: "record/upload MORE for Fish" advice is wrong beyond 4.5 min — fast clones use ≤270s; extra audio is trimmed (first 265s kept). - [ ] Fish polish: tune the emotion→marker map by ear; decide DK pro (ElevenLabs PVC, still training) role vs Fish clones; consider moving stock roster if Fish keeps winning; Resemble adapter still available for a third opinion. ## Later phases (from BUILD-PLAN.md) - [ ] Phase 2 — Stripe checkout + credit purchase (currently credits are granted by admin). - [ ] Phase 4 — actor clones, onboarding, payouts. - [ ] Password reset (email).