Public alpha · Open research release · Local-first · Open-source · Under active development
GERM is a local cultivation environment for generative microsound. It treats sound as material that can be generated, granulated, grafted, mutated, listened to, and traced through lineage. A listening from Oída can become a prompt or source in GERM; a successful render can become a descendant in Akousmata and return to Oída for another listening.
Current release: 0.5.0.
GERM is part of two Sonic Field Labs constellations: it cultivates within The Listening Stack and describes its material through the Sonic Matter Framework, whose protocol boundary is MASA 0.2.0.
GERM is an independent Sonic Field Labs project. It can use Stable Audio 3 providers, but it is not an official Stability AI product.
The Listening Stack remains model-agnostic at its contracts. GERM's first fully developed local generation path is more specific: it is designed around Stable Audio 3 and uses its text-to-audio, audio-to-audio, inpainting, continuation, variable-length, and LoRA primitives as material for cultivation. The upstream code is MIT; the currently released local weights are gated and carry separate Stability AI and component terms.
The default mock provider exercises the dashboard, graph, jobs, metadata, lineage, and WAV-writing path without downloading a model:
uv sync --extra dev
./launch_germ.commandOpen http://127.0.0.1:5178/dashboard. The mock provider emits a test signal,
not model-generated sound, so it is safe for installation and integration
checks but not a creative quality demonstration.
- Generate or transform sound through mock, local Stable Audio 3 Python, Apple Silicon MLX, or opt-in Stability API providers.
- Build a modular graph across grains, cells, swarms, tissues, controls, effects, and a realtime Chamber.
- Use local Cosmoaudition observations as attributed modulation and event material through explicit mappings, archives, and missing-data policies.
- Measure spectral, temporal, spatial, and morphological properties with Matter Analysis while keeping measurements, inference, and unavailable states distinct.
- Graft, inpaint, continue, compare, mutate, record, and collect sound while retaining prompts, seeds, models, parameters, parents, and operations.
- Import an Oída listening as sound, prompt, or lineage and ask Oída to listen to a cultivated result again.
- Write a successful generation to the shared Akousmata store only when requested.
| Path | Operating system | Hardware and software |
|---|---|---|
| Mock service | macOS or Linux | Python 3.10+, uv; CPU only; no weights |
| Stable Audio 3 Small, MLX | Apple Silicon macOS | Official sa3 optimized tools and separately accepted Small SFX or Small Music weights |
| Stable Audio 3 Small, Python | macOS or Linux | Upstream CPU path or compatible acceleration; 8 GB minimum and 16 GB system RAM suggested |
| Stable Audio 3 Medium, Python | Linux/NVIDIA | Compatible CUDA GPU; upstream reports about 6.52 GB peak VRAM, while GERM suggests 24 GB system RAM |
| Stability API | Any service-supported system | Network access, an operator-owned account, API key, and acceptance of service terms |
| Native shell | macOS 13+ | Swift 5.9/Xcode command-line tools; the Python service still performs all audio work |
Model access, RAM or VRAM use, and generation time vary substantially by checkpoint and provider. GERM reports provider diagnostics rather than promising one universal hardware minimum.
| Question | Where to begin |
|---|---|
| How does it work? | Micro/Matter architecture and provider design |
| How does it connect? | The Listening Stack and Oída integration |
| How do I install a tested model? | Local model setup |
| Which models and licenses apply? | Models and licensing |
| What is unfinished? | Known limitations and roadmap |
| How can I help? | Contribution guide |
| How should I cite it? | CITATION.cff |
- Text-to-audio generation, seed/batch variation, audio-to-audio grafting, inpainting, continuation, layer comparison, and source handoff.
- Mock, Stable Audio 3 Python, Stable Audio 3 MLX, and Stability API provider routes with readiness diagnostics and cancellable jobs.
- A browser dashboard and native macOS shell over the same FastAPI server and server-owned session graph.
- Realtime Chamber audio with granular and envelope-follower AudioWorklets, pooled trigger voices, smoothed master gain/FX updates, headroom management, compression, soft clipping, and WAV recording.
- Micro/Matter modules for grains, cells, swarms, membranes, spectral tissue, quanta, microscope analysis, saved matter profiles, biomes, and incubated evolution.
- A separate Cosmoaudition module category with cosmic, Earth, hydrosphere, biosphere, human-machine, relational, event, semantic, uncertainty, archive, mapping, and matter-processing modules.
- Optional MASA 0.2 sidecars for successful generations and Matter Analysis
artifacts. Sonic Lineage
sound_idremains GERM's canonical identity. - Wavetable Forge conversion, prompt, mutation, render, import/export, and audition routes.
- LoRA loading and a persistent Strain registry for adapter identity, intensity, tags, licensing, and provenance.
- Audio-to-control analysis, bounded OSC, MIDI intent, CV-safe exports, and a norns/Fates bridge.
- Local library, Herbarium, waveform player, metadata preview, path-safe file operations, and generation lineage.
- OÍDA re-listening and prompt derivation. GERM performs bounded local signal checks; OÍDA owns machine listening and audio understanding.
| Scale | Unit | Representative modules |
|---|---|---|
| Micro | grain, cell, quanta | Grain Culture, Cell Splitter, Quanta, Microscope |
| Meso | colony, tissue, swarm | Particle Engine, Swarm, Colony, Spectral Tissue |
| Macro | organism, culture, graph | Chamber, Genetic Matrix, controllers, performance |
All three scales use the same module graph, semantic FX bridge, sessions, library, and lineage model.
GERM belongs to two constellations. The Listening Stack is how a sound is heard, remembered, and re-listened; the Sonic Matter Framework joins three distinct projects around sound described as matter: MASA supplies the descriptive protocol, Cosmoaudition supplies attributed observations and modulation decisions, and GERM cultivates sound while retaining provenance and lineage. GERM keeps those roles distinct while making their contracts usable inside its cultivation graph:
| Component | Version / contract | Boundary in GERM |
|---|---|---|
| MASA | 0.2.0 (MIT) | Optional descriptive JSON sidecars under output/masa/, citing the published canonical schema at masa.sonicfield.org. They never replace Sonic Lineage or turn a successful render into a failure. |
| MASA Processing | masa-processing-request 0.2.0 |
Every Micro module declares a granular or spectral operation in MASA's engine-neutral terms. GERM states the intention and binds no DSP library. |
| GERM Matter Analysis | — | A bounded local analyzer informed by the shared measured / inferred / unavailable distinction; it is not a claim of listening or a fourth framework project. |
| Cosmoaudition | cosmo/modulation/v0.2 and cosmo/signal-catalog/v0.2 |
A loopback-only, response-bounded HTTP bridge reads source status, canonical signal definitions, snapshots, and modulation frames. GERM retains v0.1 frame compatibility and never contacts observation providers directly. |
MASA 0.2.0 carries Roads's granular vocabulary as a protocol layer, so a Micro module can say what it wants done to matter without naming an engine:
| Micro module | Operation | Character |
|---|---|---|
| Grain Culture | matter.granulate |
Quasi-synchronous emission at the perceptible grain. |
| Particle Engine | matter.granulate |
Dense asynchronous emission, expodec envelopes. |
| Quanta | matter.granulate |
The millisecond threshold where a grain stops being a small note and becomes a particle. |
| Cell Splitter / Colony | matter.fragment |
Transient-led and grain-led division. |
| Spectral Tissue / Membrane | matter.extract |
Spectral strata and bands. |
| Microscope | matter.reduce |
Spectral peaks retained for inspection. |
| Metabolism | matter.timestretch |
Duration change with declared transient handling. |
| Swarm | matter.pitchshift |
Transposition with declared formant handling. |
GET /micro/processing-operations reports the mapping; POST /micro/processing-request builds a portable request. A request is an
intention, not a receipt: it asserts nothing about what was rendered or heard.
Cosmoaudition mappings are operator-authored control relations. They do not claim that a dataset is the literal voice or identity of a source. Missing or unfetched observations do not silently become zero or a neutral modulation; the route remains inactive until its state is explicit.
GET /cosmoaudition/frame reads one modulation frame and resolves it into
GERM routes. It reads the frame's controls, never its bare values map:
a value arrives with the decision that produced it, or it does not arrive.
Out-of-range or structurally invalid controls are withheld with an explicit
bridge reason. Frame-level signals and sources travel beside the routes so
their source, unit, sphere, epistemic status, temporal character, signal kind,
and declared normalization remain
traceable through signalId. Withheld routes and emitted absences are reported
rather than dropped. Server-Sent Events are deliberately not bridged, because
this bridge is a bounded request/response client; poll /cosmoaudition/frame
instead.
GET /cosmoaudition/signals exposes the source-filterable catalog used by
Cosmoaudition v0.2. GERM validates its finite input ranges, [0, 1] output,
clipping rule, attribution fields, and epistemic axes before using it. The
Hydrosphere Field selects marine and hydrological observations without folding
them into the atmospheric/geological Earth field.
Oída hears. GERM cultivates. Akousmata remembers. AKOÚŌ structures. Earworm routes. Together they form The Listening Stack, open infrastructure for listening, re-listening, sonic memory, and cultivation.
| Component | Version / contract | GERM integration |
|---|---|---|
| OÍDA | 0.10.0 / oida/gateway/v0.6 |
Re-listen to generated sound, derive editable prompts, and retain a listening only when requested. |
| Earworm | 0.7.0 / Akousma spec 1.6 / earworm/auditum/v2 |
Preserve accountable listener types, route decisions, honest absence, disagreement, provenance, lineage, and additive revisions without treating them as equivalent. |
| Akousmata | 0.7.0 | Import remembered sound, prompt, or lineage; write successful generations back as child akousmata and retain supplied auditum blocks. |
| AKOÚŌ | 0.9.2 / akouo/v0.9 |
Keeps embodied reports distinct from machine inference while structuring evidence permissions, apparatus, temporal passes, and covenants. |
| Algophony | 0.5.2 | Can evaluate lineage-bearing generation batches without changing GERM's generation state. |
| ORAM | 0.4.1 | Uses the local GERM-compatible generation surface for constrained sound summoning and transformation. |
The core handoff is:
Sound → Oída hears → GERM cultivates → Akousmata remembers
↑ | |
+------------- listen again -----------+
AKOÚŌ structures claims and routes. Earworm routes events,
provenance, retention, and lineage between the organs.
For a guided GERM or complete Listening Stack installation, including model choice, storage and memory guidance, gated-access checks, downloads, and local gateway configuration:
curl -fsSL https://raw.githubusercontent.com/sonicfieldlabs/listening-stack/main/install.sh | bashChoose GERM only or Oída + GERM in the terminal assistant. GERM remains in this repository; the installer only coordinates its source, dependencies, provider, models, and local configuration.
Requirements: Python 3.10+ and uv.
uv sync --extra devInstall a local Stable Audio provider only when needed:
uv sync --extra python-provider
./scripts/install_mlx_provider.sh # Apple Silicon routeProvider model repositories and weights remain outside Git. See local setup and provider design.
./launch_germ.commandOpen http://127.0.0.1:5178/dashboard.
Server only:
./scripts/run_server.shNative macOS shell:
apps/macos/script/build_and_run.shThe shell embeds the same dashboard and supervises the same daemon. It does not maintain a second session or generation state.
Health and diagnostics:
curl http://127.0.0.1:5178/health
curl http://127.0.0.1:5178/diagnosticsThe complete request and response contract is in docs/api_reference.md. Principal route groups:
| Group | Routes |
|---|---|
| Generation | /generate, /audio-to-audio, /inpaint, /continue, /jobs/* |
| Models | /models, /models/load, /diagnostics, /huggingface/status |
| Listening | /listener/enhance, /listener/score, /listener/relisten |
| Memory | /earworm/export, /import, /akousma/* |
| Micro/Matter | /micro/matter-profile, /matter/analyze, /micro/matter-analysis, /micro/biomes/* |
| Cosmoaudition | /cosmoaudition/status, /cosmoaudition/modules, /cosmoaudition/sources, /cosmoaudition/snapshot, /cosmoaudition/map, /cosmoaudition/archives/* |
| Strains | /strains/*, /lora/load, /lora/strength |
| Wavetables | /wavetables/* |
| Control | /control/* |
| Sessions | /sessions, /sessions/current, /sessions/{id} |
| Files/library | /library, /files/* |
Example mock generation:
curl -X POST http://127.0.0.1:5178/generate \
-H "content-type: application/json" \
-d '{
"provider": "mock",
"model": "mock-sine",
"prompt": "dry close-microphone loop of metallic friction",
"negative_prompt": "speech, vocals, melody, long reverb",
"duration": 3.0,
"steps": 8,
"cfg_scale": 1.0,
"seed": -1,
"output_name": "metal_friction_sprout",
"tags": ["SFX", "loop", "texture"]
}'- The default provider is
mock; no model or weight is downloaded implicitly. - Local input, output, model, upload, and duration limits are configured in
.env.example. - Hosted generation and cloud vision remain opt-in. Credentials are read from the process environment and are never stored in generation metadata.
- The server binds to loopback and validates Host/Origin headers by default.
- Generated audio, metadata, uploads, sessions, strains, model weights, build products, and local app state are ignored by Git.
- A generation written to Akousmata records only portable identifiers and relative/protocol-safe provenance. Machine-specific paths stay local.
- This is a public alpha. Session, graph, and provider interfaces can still change before 1.0 even though public contracts are documented.
- Jobs are held in memory. A server restart loses job state, while completed WAV and metadata files remain on disk.
- A queued Python-provider job can be cancelled, but an in-process render may finish because the upstream API has no safe mid-render interrupt. MLX subprocess jobs support active cancellation.
- Real-model access is separately licensed and may be gated. Output quality, latency, determinism, and hardware use vary by provider and checkpoint.
- The native macOS shell is built locally and is unsigned by default.
- Physical CV output is disabled. MIDI may use browser Web MIDI or an optional configured local backend; otherwise the server records intent only.
- Hosted routes are explicit opt-ins and can send prompts, source audio, or derived material to their provider. The default mock route stays local.
- GERM performs bounded signal checks but does not replace Oída's listening and claim-accountability layer.
- Cosmoaudition live data requires the separate local Cosmoaudition System; fixture mode and archived observations remain available without granting GERM direct provider access.
The roadmap identifies current research priorities and non-goals.
uv run pytest -q
uv run ruff check server tests
node --check dashboard/static/app.js
node --check dashboard/static/dish.js
node scripts/smoke_dashboard.mjs- API reference
- Architecture of Micro/Matter
- Stable Audio integration
- OÍDA and Akousmata integration
- Provider design
- Local setup
- Native macOS shell
- Troubleshooting
- Models and licensing
- Public roadmap
- Changelog
MPL-2.0. See LICENSE.