Psychology VR - MetaHuman cast, approved voices and diagnostics

v0.5.0 — MetaHuman cast, approved voices and diagnostics

This is a complete Windows application ZIP. Extract the whole PsychologyVR
folder; no previous release or patch folders are required. Initial setup still
downloads the local AI models. Unity Editor and Unreal Engine are not needed to play.

Included

• Four clothed MetaHumans in the consultation game: Alex/Dominic, Ivan/Bruce,
  Stefan/Mateo and Nikola/Victor. Baked facial reactions, neck movement, tears
  and speech/emotion blending are integrated. Torso/hand animation is deferred.
• Scenario voice assignments: Igor for burnout, Person 1 for fire, Person 2 for
  bereavement and Person 3 for earthquake. Runtime packs contain the selected
  references, transcripts, hashes and emotional synthesis settings. Person 1
  uses the approved mono references with background noise and reverb reduced.
• BF16 Higgs quality retained. Whole emotional beats keep their approved delivery
  instead of being forcibly split into sentences. Presets specifically approved
  with sentence pauses keep those pauses. Interrupted unfinished speech is not
  assumed to have been heard in full.
• Prepared voice references and cached fixed opening lines; fresh replies still
  require generation. No student conversations are cached as stock answers.
• Gemma remains the default. Patient prompts distinguish personal experience
  from expert explanation; simple questions no longer trigger a forced minimum
  length rewrite. Requested detail remains possible.
• Desktop borderless fullscreen/windowed settings and adjustable larger subtitles.
• Automatic resource samples every two seconds: GPU utilization, dedicated VRAM,
  GPU temperature and system RAM. Export Diagnostics.cmd creates a local ZIP
  of the last seven days of resources and reply timings, with a readable summary.
  Cached opening lines are reported separately. Dialogue, recordings, credentials
  and raw service logs are excluded; there is no automatic upload.
• Existing text conversation tester, automatic session journals, offline voice
  previews and standalone facial/head-motion preview tools are retained.

Upgrade

Close the game and run Stop Services.cmd. Merge the complete ZIP's
PsychologyVR contents into your existing installation, replacing matching files.
Keep .cache, services/config.local.json and all logs/. Use the normal
Start Desktop.cmd or Start VR.cmd launcher afterward. A full service
restart is necessary to load the updated speech, dialogue and diagnostics code.

Fresh installations default to Gemma 4 12B QAT Q4, BF16 Higgs and CPU Whisper.
Existing configuration is preserved; use PC Settings.cmd if it selects an
older model. The release does not bundle model weights, local configuration,
student sessions, raw face-capture footage, Unreal builds or editor caches.

Known limitations and verification

Gemma can still produce malformed regional-language words and awkward grammar.
The latest prompt/length changes are a partial improvement, not a fluency fix.
Person 1 and Person 2 have no approved sadness/happiness runtime recipes yet;
these emotions retain that speaker's neutral voice rather than inventing a preset.
Headset stability, VR frame rate and RTX 4090 performance still need university
testing. GPU measurements cover the whole PC, and two-second samples can miss
shorter peaks. First-audio readiness excludes microphone/STT time and does not
prove the listener heard the audio.

The existing final Unity Windows build passed before packaging. The full Python
service suite passes 165 tests. The diagnostics recorder was exercised against
the local NVIDIA GPU and its ZIP exporter. Package validation checks required
assets, offline voice references and archive CRCs. The release manifest records
the source commit and file hashes; SHA256SUMS covers the ZIP and manifest.

Psychology VR - Voice library and animation tools

Complete Windows package with the new voice library and acted-animation tools, including all previous Gemma Unity updates.

Included

• All three contributed speaker packs: original recordings, prepared audio, transcripts and selected voice presets.
• Person 3's approved anger (bitterness), sadness (restrained delivery), and happiness (longer happy reference + enthusiasm).
• Preview Voices.cmd, an offline listening page with generated examples and their recorded references. No AI setup is needed to listen.
• Rebuilt Unity consultation player and separate facial/head-motion preview player. Drag a prepared take folder onto Preview Recorded Face.cmd; personal capture videos and extracted take data are excluded.
• Voice preparation/audition scripts, MediaPipe extraction tools, the Montenegrin project report, and the requested statistics-line sample in WAV/MP3.
• Updated setup, repository and release documentation. The Unreal prototype is available in the source repository; an Unreal runtime is not included in the Windows ZIP.

Current scope

The new speaker packs are ready for development and listening, but are not automatically assigned to individual patients. The game continues using its configured shared voice. The recorded-animation player is a separate proof of concept; live consultation animation integration, further rotoscoping, longer VR stability/stress tests and RTX 4090 acceptance are still pending. This release does not claim to fix the previously observed Quest Link freeze.

Validation

• 136 service tests and all 4 recorded-head-pose tests passed; GitHub CI passed.
• Both Unity Windows builds succeeded. The facial bake matched 222 mesh bindings with zero sampled facial error and zero head-rotation error.
• The rebuilt consultation player passed pause/resume and interruption through an isolated HTTP/audio fixture. This check used generated test tones, not live AI or a headset.
• ZIP CRC, all 734 packaged file hashes, offline listening-page links and person 3's preferred sample/reference hashes verified.
• Fresh extracted-package configuration selected Gemma + BF16 Higgs + CPU STT with simulated 24 GB GPU metadata. This is a setup check, not an RTX 4090 performance test.

Download and update

Download psychology-vr-windows.zip and extract the complete PsychologyVR folder. Use Start Desktop.cmd, Start VR.cmd, Start Text Chat.cmd, or Preview Voices.cmd. First AI setup downloads the pinned models and runtimes.

For upgrades, stop the old game and services, then merge this complete package into the existing installation. Preserve .cache, services/config.local.json and logs/sessions. No patch ZIP or earlier release is required. GitHub's Source code archives are for development and do not contain the built player.

Voice assets are included as requested by the project owner. Credentials, local configuration, model weights, personal facial-capture data and student session logs are excluded. SHA-256 checksums and a per-file manifest accompany the ZIP.

Psychology VR - Gemma Unity complete package

Complete Unity Windows package with Gemma 4 12B QAT Q4 as the default dialogue model, retaining BF16 (16-bit) Higgs expressive speech and CPU speech recognition.

• Adds Settings → Dialogue model for switching between installed Gemma, Bonsai and original Qwen alternatives between replies. PC Settings options 5 and 6 install Bonsai and Gemma.
• Includes the built Unity player, all previous application updates, setup tools, room and character assets, and the owner's approved voice reference. Unreal is excluded.
• The source repository now includes the multi-turn model evaluation runner, retained optional Model Lab UI, comparison results and GPU-memory reports.
• Validation: 136 service tests passed; Unity Windows build succeeded; archive CRC and every packaged file hash verified; fresh extracted-package configuration selects Gemma + BF16 on a simulated 24 GB GPU configuration. This setup check does not simulate runtime GPU performance.
• Real Quest Link memory measurement peaked at 22.40 GiB total on an RTX 5090. A headset freeze was reported while desktop playback continued. The user attributes it to an NVIDIA driver issue; this has not been independently confirmed. A longer headset test and testing on the university RTX 4090 remain necessary.

Download psychology-vr-windows.zip, extract the complete PsychologyVR folder, then run Start VR.cmd, Start Desktop.cmd or Start Text Chat.cmd. First setup downloads the selected AI model and speech dependencies; subsequent launches are local. Do not use GitHub's Source code ZIP to run the player.

For upgrades, stop the old game and services, then merge this complete package into the existing installation. Keep .cache, services/config.local.json and logs/sessions. If a previously saved explicit model/preset remains selected, use PC Settings.cmd → 6 Gemma. See START HERE.md. No patch folders or previous release downloads are needed.

The reference voice is included with its owner's authorization for this private release. Student session logs and machine-specific configuration are excluded.

Psychology VR - Complete Windows package

Download psychology-vr-windows.zip below. This is the complete application: game, room and character assets, AI service code, launchers, approved reference voice and every previous fix. You do not need any older release or patch ZIP.

Extract the full PsychologyVR folder and run Start VR.cmd, Start Desktop.cmd or Start Text Chat.cmd. First launch installs the private runtime and downloads the selected AI models automatically. Unity and a separate Python installation are not required. Allow approximately 45 GB free space and internet for first setup.

Upgrading an existing copy: close the game and run Stop Services.cmd first. Extract this complete ZIP into the parent of your existing PsychologyVR folder and replace matching files. Keep the existing folder so its model downloads, local configuration and session logs are retained. If you renamed the installation folder, copy all contents of the ZIP's PsychologyVR folder into it. START HERE.md includes the steps.

This release changes packaging and instructions only. It retains all v0.2.3 features, including PC Settings.cmd → 4 Fast dialogue, the original dialogue presets, automatic GPU detection and the approved 16-bit/BF16 voice. Fast dialogue is optional: the smaller model has known language and character-consistency weaknesses. It reduces dialogue delay but does not eliminate pauses caused by speech synthesis. The measured 13.1-second first speech was on an RTX 5090, not a university RTX 4090.

Requires Windows 10/11 x64, current NVIDIA drivers, one NVIDIA GPU with at least 24 GB VRAM and preferably 64 GB RAM. VR uses Meta Quest Link.

The reference recording is included with the owner's permission in this private release. Session logs and machine-local settings are excluded. The AI weights are downloaded by Setup rather than bundled in the ZIP. The other two assets are integrity records; GitHub's automatically generated Source code downloads are for development.

Psychology VR - Fast dialogue trial

Adds an opt-in Fast dialogue setting for the long waits reported with the 27B CPU-offloaded model. Existing Auto and 27B presets are unchanged.

Try it on an existing installation

Download pc-settings-update.zip, close the game, and extract the update into your existing application folder, replacing matching files. Run PC Settings.cmd and choose 4 Fast dialogue. It downloads the pinned Qwen3.5 9B Q6 model (approximately 7.4 GB) and runs its layers entirely on the GPU. Keep .cache, .tools, voices and logs. The original models remain available for switching back.

BF16 speech quality, the approved voice reference, authored character profiles, emotion controls, relationship policy, memory and session logging are retained. The backend also improves length retries: progressively shorter rewrite targets replace repeated identical requests after an overlong response; it does not truncate spoken text.

What was measured

On the local RTX 5090, four same-question dialogue-only probes took 1.53-1.72 seconds, compared with 26.54-45.46 seconds for the earlier IQ3_M / 32-layer configuration. Wording differs between models; this is a small application screen, not a general benchmark.

A complete built-player run passed speech and lip movement:

• Dialogue/appraisal preparation: 2.494 s.
• First audible speech: 13.131 s (the previous 27B observation was 67.224 s with different wording).
• All three sentences prepared: 30.944 s, with later synthesis overlapping playback.
• BF16 synthesis across the reply: 25.446 s; lip alignment: 2.985 s.
• Total GPU peak, including other applications: 21,421 MiB / 20.9 GiB.

This is not a 4090 headset benchmark or a promise of those times on the university PCs. BF16 speech synthesis is now the main measured delay. All response audio being ready is different from the time the character starts or finishes speaking. Inter-sentence gaps remain: synthesis already overlaps playback, but the tested worker produces audio more slowly than it plays (about 6 seconds of starvation between the first two test sentences). Fast reduces dialogue latency; it does not fix BF16 speech throughput.

Quality tradeoff

Fast remains a trial option. Samples show grammar, gender agreement and dialect errors, awkward phrasing, and an inappropriate implication of guilt in a bereavement response. It is not quality-equivalent to the larger model. Review text conversations with colleagues before using it with students. Auto continues using the 27B presets.

111 service/helper tests pass. A seven-turn conversation-policy screen completed after the retry fix, including boundary escalation and session exit. Contract success does not establish clinical or native-language quality. See docs/GPU_PRESETS.md and the source validation records for details.

For a new PC, use psychology-vr-windows.zip, which contains the game and the owner's approved reference voice in this private release. The small update contains neither the voice nor the game. Requires Windows 10/11 x64, current NVIDIA drivers, a single 24 GB+ NVIDIA GPU, and preferably 64 GB RAM. Quest VR requires Meta Quest Link. Allow about 45 GB free disk space. Source code ZIP is for development.

Session logs and machine-local configuration are excluded. SHA256SUMS.txt and release-manifest.json record the exact source commit and asset integrity.

Psychology VR - Automatic GPU presets

Adds PC Settings.cmd with Auto-detect GPU, University / RTX 4090, and Original quality options.

Auto selects IQ3_M dialogue with 32 GPU layers on a 24 GB card, or the existing IQ4_XS / 48-layer setup on a 32 GB+ card. BF16 speech, your reference voice, character prompts and session logs are unchanged. These choices use the capacity of a single CUDA device, not combined memory across PCs.

Existing installation

Download pc-settings-update.zip, close the game, and extract the patch into your existing application root, replacing matching files. Double-click PC Settings.cmd and choose 1 (Auto). Setup stops this installation's AI services, backs up configuration changes, and downloads only the selected dialogue quant. The IQ3_M model is a one-time 12.8 GB download. Keep .cache, .tools, voices and logs; existing models and shared speech/STT files are reused. Start VR or Desktop normally after setup.

New university PC

Download psychology-vr-windows.zip, extract the complete PsychologyVR folder, and use a Start launcher. Auto is the default. Requires Windows 10/11 x64, current NVIDIA drivers, a single 24 GB+ NVIDIA GPU and preferably 64 GB RAM. Quest 3 uses Meta Quest Link. Allow about 45 GB free space. Initial model downloads are approximately 25 GB for the 4090 preset or 28 GB for original quality, plus runtime dependencies. No Unity/Python/Git installation is required. The full private release includes the owner's approved voice; the small patch does not include the voice or game.

Validation and tradeoff

110 tests pass. The real settings menu and an extracted installation were tested. Full dialogue, BF16 playback and lip movement passed with the 32-layer preset on an RTX 5090. Peak total GPU use was 21,299 MiB (20.8 GiB), including other GPU applications on that host. First speech took 67.224 seconds in that particular run: CPU offload saves GPU memory but can increase waiting. This is not a speed improvement claim.

The three university RTX 4090 / 64 GB PCs still need a Quest Link acceptance session; no actual 4090 or headset was available for this release test. See docs/GPU_PRESETS.md for the test scope and colleague review checklist. Original quality remains selectable. A smaller-parameter dialogue model is the fallback if the 4090 preset remains too slow or memory constrained.

Checksums and the exact source commit are recorded in SHA256SUMS.txt and release-manifest.json. Session journals and local configuration are excluded from release assets.

Psychology VR - Setup configuration fix

Fixes KeyError: 'higgs_reference' when Setup encounters an older or partial services/config.local.json.

Repair an existing installation

1. Download setup-fix.zip (the small patch).
2. Extract it into your existing psychology-vr application folder. Allow it to replace tools/setup-portable.py.
3. Double-click Setup.cmd again.

Keep your .tools, .cache, voices and logs folders. Existing dependencies and model downloads are reused. Setup adds missing defaults, preserves explicit preferences and makes a byte-for-byte backup of an upgraded configuration under services/.runtime/config-backups/. It now reports invalid configuration and missing reference files clearly.

New installation

Download psychology-vr-windows.zip, extract the complete PsychologyVR folder, and use Start Desktop.cmd, Start VR.cmd or Start Text Chat.cmd. This full ZIP includes the game and the owner's approved voice in this private release. GitHub's source ZIP does not include them.

First launch automatically installs Python and dependencies and downloads approximately 28 GB of models. Allow about 45 GB free disk space. Requires Windows 10/11 x64, a current NVIDIA driver and at least 24 GB VRAM; 64 GB system RAM is recommended. Quest VR requires Meta Quest Link. Speech remains BF16. See START HERE.md for controls and system prerequisites.

106 tests pass, including four configuration regression cases. The extracted installation was checked with portable configuration keys deliberately removed; migration and the installed runtime check passed (PORTABLE_READY). Full game/speech behavior is unchanged from v0.2.0; no new headset acceptance test was performed for this installer-only fix.

Session logs and local configuration are excluded from all downloads. The full release includes the approved voice; the small patch contains only the corrected setup helper and instructions. SHA256SUMS.txt and release-manifest.json record package integrity and the exact source commit.

Psychology VR - Windows testing build

Download psychology-vr-windows.zip, extract the complete PsychologyVR folder, and double-click Start Desktop.cmd, Start VR.cmd, or Start Text Chat.cmd. Do not use GitHub's Source code ZIP to run the game.

The first launch automatically installs its folder-local Python runtime, speech dependencies, llama.cpp, and pinned AI models. No Unity, Git, Python or paid API setup is needed. Allow internet access, approximately 28 GB of model downloads plus runtime downloads, and about 45 GB free disk space. Later launches use local files.

Requires Windows 10/11 x64, a current NVIDIA driver and at least 24 GB VRAM; 64 GB RAM is recommended. For Quest 3, install Meta Quest Link and enter PC Link before launching VR. The Windows Visual C++ runtime is also a system prerequisite; START HERE.md links the installer if needed. This is PCVR, not a standalone Quest APK. Validation used an RTX 5090; the 4090 target still needs its own hardware acceptance test.

Included changes:

• Four adult patient cases, improved conversational flexibility, bounded memory and relationship/boundary responses.
• Automatic local session journals for text and VR conversations.
• Sentence-by-sentence BF16 speech playback, interruption-aware history and refined emotional delivery.
• Improved Montenegrin recognition and Latin-script output.
• Complete built game, room/character assets, automatic setup, repair and stop launchers.
• The owner's approved reference recording and transcript, included with explicit permission in this private release only. Keep them within the intended testing group.

Verification: 102 service/release tests pass. An extracted folder with spaces installed a fresh Python environment and dependencies, verified pinned model hashes, launched the game and text tester, and passed streamed pause/interruption and full Qwen/BF16 playback checks. Large model weights were reused from a verified cache for that installation test. This was not a clean Windows VM or a new headset acceptance test.

Sessions stay in logs/sessions/ and are not uploaded. Keep this folder when upgrading; stop services from the previous installation before running a different copy. Read START HERE.md for controls and troubleshooting.

SHA256SUMS.txt and release-manifest.json provide archive integrity and the exact source commit/file inventory. Model downloads retain their upstream terms, including Higgs's research/non-commercial terms.

Psychology VR prototype — Windows builds and source assets

Existing Windows development builds for the Psychology VR prototype.

Downloads

• psychology-vr-windows.zip — consultation game; player data last updated 15 September 2026.
• psychology-vr-character-preview.zip — older jumper character audition from 11 September 2026.
• SHA256SUMS.txt and build-manifest.json — archive checksums and per-file hashes.

Extract a ZIP into your repository root. It restores unity/Builds/Windows/ or unity/Builds/CandidatePreview/, including the executable, data and required DLLs. These archive the existing local builds; they were not rebuilt from the release's packaging commit. Unity debugging backup folders are omitted.

Running the consultation game

Follow the README's Python setup. tools/launch.ps1 -Scripted -Desktop runs with scripted dialogue and Windows speech, without downloaded AI weights. For live conversations, follow the full model/runtime setup, copy the live configuration template, then run tools/launch.ps1 -Desktop (or omit -Desktop for Quest Link). Kokoro is still the game's TTS; Higgs remains an audition candidate.

The archives do not bundle model weights, Python environments, personal voice recordings or cloned speech. The separate character-preview build is an audition tool, not the full consultation game.

Source assets

The repository now also includes the original man/, jumper_man/, and Characters_with_expressions/ packs through Git LFS, in addition to converted Unity assets. Clone with Git LFS to retrieve these. Third-party asset credits and terms remain applicable.