Visual architecture and work breakdown for Project Talamaska. Epics 1–5 implement the symbolic-first solo-piano pipeline; Epics 6–8 cover data, evaluation and UX. This page mirrors the EPIC/STORY map used in the GitHub project.
Text-to-piano pipeline with symbolic control
(MidiLikeScore) before audio. Epics 1–5 map directly onto
these five stages.
Raw Text
User prompt captured via web UI and passed into the orchestration layer. Linked to Epic 8 (UX).
Text → ControlDict
Rule/LLM-based prompt conditioning mapping text to
ControlDict. Backed by
Epic 2 (Prompt Controls).
Symbolic Generation
Conditional Transformer generating
MidiLikeScore from controls + optional prefix. Backed
by Epic 3 (Composer Model).
Performance Layer
Converts quantized score to expressive
MidiLikePerformance controlled via a single slider.
Backed by Epic 4 (Humanizer).
Audio / Export
Sample-based piano rendering to WAV/FLAC plus MIDI export. Backed by Epic 5 (Renderer & Export).
The five core architectural blocks and how they tie back to epics and stories.
Epic 1 – Pipeline Skeleton & APIs
Defines MidiLikeScore and
MidiLikePerformance, shared module APIs and the
end-to-end orchestrator. Provides caching and partial regeneration
hooks.
Epic 2 – Prompt Controls & Encoding
Maps free text to ControlDict and control tokens
(mood, tempo, meter, density, length, key). Includes schema,
rule-based parser and metadata → labels tooling.
Parameter extraction weights · E2-01/02/03
Epic 3 – Symbolic Composer Model
Generates MidiLikeScore: the “sheet music” layer.
Uses event-based token vocabulary and a conditional Transformer,
trained on clean solo-piano datasets.
Epic 4 – Performance Humanization
Applies timing jitter, velocity shaping, pedal and phrase
dynamics, controlled by PerformanceSettings and a
single “humanization” slider.
Quantized grid vs. humanized performance · E4-01/02/03
Epic 5 – Renderer & Export
Renders MidiLikePerformance into WAV/FLAC and MIDI
with loudness normalization and clipping checks.
Relative effort in weeks for Epics 1–8. Phase 1 focuses on Epics 1–5 (walking skeleton); Phase 2 adds Epics 6–8 (data, evaluation, UX).
Target processing time allocation per request (Total: < 10s). Segments map directly to epics: Parser (E2), Composer (E3), Humanizer (E4), Renderer (E5), Overhead (E1 + E8).
Eight epics grouped around architecture blocks and supporting capabilities. Each card lists the main outcomes and key stories.
End-to-end Python skeleton, symbolic types and shared API contracts for all modules, including caching and partial regeneration hooks.
Map free-text prompts to a well-typed ControlDict and
control tokens usable in both training and inference.
Conditional Transformer that generates MidiLikeScore
from controls and optional prefix score.
Deterministic humanization layer that turns quantized scores into expressive performances with a single “humanization” slider.
Stable, repeatable rendering from performance events to audio and MIDI with quality checks and simple export API.
Clean, reproducible pipeline from raw piano datasets to tokenized train/val/test sets and training jobs.
Symbolic and audio metrics, plus prompt controllability and humanization A/B experiments.
Minimal but robust web shell that exposes prompt → audio, partial regeneration and clip telemetry for future governance.
Sunburst view of the EPIC/STORY map used in the GitHub project. Center: Project Talamaska. Ring 1: Epics 1–8. Ring 2: key stories (E*-0*).
Epic 4 maps ControlDict and slider values to concrete
performance parameters. This stays explainable for creators.
Comparison of performance parameters for two prompts, using the
same model but different ControlDict and slider inputs.