ZPE-Prosody

Package Install

Installable package: python3.11 -m pip install zpe-prosody. Current release: 0.1.1 on PyPI. Source: Zer0pa/ZPE-Prosody.

python3.11 -m pip install zpe-prosody

For full install, smoke, source, and developer commands, click here.

---

00 ZPE-PROSODY · CONTOUR CODEC

RESEARCH-READY · RETRIEVAL OPEN

Encoding Speech's Shape And Feeling

Live experiment. Not a release. Ambition headlined. Claims bounded.

Prosodic feature-store codec · pitch, energy, duration, voiced mask · PyPI zpe-prosody v0.1.1 is stale · github.com/Zer0pa/ZPE-Prosody

A voice carries more than words. Pitch rises into a question, stress lands on a syllable, and rhythm tells a listener whether the speaker is calm or in a hurry. ZPE-Prosody captures that shape as a deterministic ZPRS/v1 stream at 13.0× mean compression and 0.64% voiced-F0 RMSE on 100 LibriSpeech test-clean utterances.

It stores acoustic prosody cues, not emotion, intent, semantic meaning, or a speaker-state diagnosis.

The encoder is the product. Retrieval misses target. Transfer is paused. Both limits are named here instead of hidden behind the compression win.

Scope: encoder stream for F0, energy, duration, and voiced mask. Retrieval misses target; transfer remains paused.

---

02 Markets

Adjacent forecasts, not product claims

Speech and language processing '30 · $26.8B    Text-to-speech market '31 · $7.9B    Text-to-speech software '30 · $7.3B    Voice analytics '30 · est. $3.1B    Speech AI / feature-store tooling '30 · est. $1.8B

ZPE-Prosody is a bounded prosodic encoder. Retrieval and transfer are not claimed.

03 Value of Market $7.3B TTS market by 2030; the prosodic feature store beneath it, with the retrieval gap stated.

04 Insight Speech carries feeling. Its shape can now be held.

---

05.1 Current Tech Computed and discarded Mainstream TTS and voice-analytics stacks repeatedly compute pitch, energy, and timing, then discard the contours or store them as undocumented bytes. Buyers get no shared archive format, no published fidelity number, and no explicit retrieval limit.

05.2 Our Tech The shape, held ZPE-Prosody stores F0, energy, duration, and voiced mask as a deterministic ZPRS/v1 stream. The public receipt is 13.0× mean compression, 0.64% voiced-F0 RMSE, 2.67 ms mean encode latency, and four primitive checks passing. Retrieval and transfer are excluded from the product on purpose, with the numbers.

05.3 Benchmarks

LibriSpeech test-clean, 100 utterances

Compression · 13.0×    F0 RMSE · 0.64%    Primitive checks · 4/4 PASS    Retrieval · MISS, p@5 0.31    Transfer · PAUSED_EXTERNAL

PRO-C006 misses target at p@5 0.31 vs 0.80. PRO-C005 transfer is paused; no commercial-safe substitute is proven in-lane.

---

06 Measurement PRO check suite Four checks pass. Retrieval and transfer do not.

06.1 Comparative Performance LibriSpeech contour compression ZPE-Prosody — 13.0× compression gzip — ~2.2× raw PRO-C006 p@5 — 0.31 MISS PRO-C004 — PASS The primitive encoder checks pass. Retrieval misses at p@5 0.31 vs 0.80; OOD p@5 is 0.1707. Transfer remains paused.

---

07 Key Metrics

LibriSpeech test-clean

07.1 F0 RMSE · 0.64% Voiced-frame error on the 100-utterance LibriSpeech test-clean receipt.

07.2 Compression · 13.0× Mean compression against raw float32 prosody contours in the deterministic ZPRS/v1 stream.

07.3 Primitive Checks · 4/4 PASS PRO-C001 through PRO-C004 pass for the primitive encoder only; this does not override retrieval or transfer gates.

07.4 Corpus · 100 utterances LibriSpeech test-clean, OpenSLR source, fixed as the public benchmark receipt for this README surface.

07.5 Retrieval Target · 0.31 p@5 MISS The retrieval gate misses against the 0.80 threshold and stays named as a blocker, not stretched into a hidden product claim.

---

08 Encoder Bounds

The encoder holds speech's shape. Retrieval does not yet follow.

08.1 What Round-Trips Exactly On 100 LibriSpeech test-clean utterances, the encoder records 13.0× mean compression at 0.64% voiced-F0 RMSE, with duration RMSE of 0.000 ms across 5/5 hash-identical encoder runs. The same input bytes produce the same ZPRS/v1 stream every time, on every host. PRO-C001..PRO-C004 pass on primitive encoder checks. They do not override the retrieval and transfer gates. The page reports both, not one.

08.2 Honest Blocker MISS on PRO-C006 retrieval, p@5 0.31 vs 0.80; OOD p@5 0.1707. PRO-C005 transfer is PAUSED_EXTERNAL; no commercial-safe substitute is proven in-lane. Status packet on PR #50 branch-public; PyPI stale at v0.1.1. No transfer-learning product, retrieval product, or TTS-ready system is claimed.

---

09 A Voice with a Fidelity Receipt

09.1 The Ambition The product is a bounded ZPRS/v1 feature store for the shape of speech: F0, energy, duration, and voiced mask. TTS teams, call-centre analytics owners, and linguistics labs can store, ship, and re-read prosody with a stated fidelity per recording. Retrieval and transfer arrive later, on their own terms.

09.2 What Works Now The prosodic encoder ships with a fidelity number per frame and a public compression figure.

09.3 What's Still Open Retrieval misses target at p@5 0.31 vs 0.80. Transfer is paused on an external dependency.

09.4 Feature Stores · Near-term, 12-24 mo TTS teams stop drowning in contour bytes. A TTS platform keeping pitch and energy contours for thousands of speaker voices and styles cuts feature-store storage by roughly 87% against its current gzip baseline. The same archive holds many more voices on the same disk.

09.5 Fidelity · Near-term, 12-24 mo Voice pipelines inherit a pitch receipt. A voice-cloning engineer who round-trips a speaker through the codec sees the F0 error per utterance, 0.64% on LibriSpeech, before the model ingests the contour. Pitch drift becomes a dashboard number, not a listener complaint.

09.6 Call Centres · Mid-term, 24-48 mo Analytics vendors archive prosody, not just transcripts. A call-centre analytics platform that already stores transcripts can store prosody beside them at a tractable cost. Emotion-AI and sentiment systems get to work from the actual shape of how a customer spoke, not a downstream summary of it.

09.7 Linguistics · Mid-term, 24-48 mo Prosody corpora become comparable. A linguistics lab studying stress and intonation across dialects can compress a multi-year recording corpus into a portable feature store with a stated pitch error. A peer can reproduce the analysis on the same bytes, not on a re-derived contour.

09.8 Disclosure · Paradigm, 48 mo+ Speech feature codecs get fidelity terms. A market in which prosodic codecs publish compression, F0 RMSE, and the retrieval limit side by side changes how buyers procure speech tooling. A TTS vendor can talk to a regulator and a customer with the same numbers, units, and corpus.

---

Install / Developer Commands Detailed

Package Install

Installable package: python3.11 -m pip install zpe-prosody. Current release: 0.1.1 on PyPI. Source: Zer0pa/ZPE-Prosody.

python3.11 -m pip install zpe-prosody

Import smoke:

python3.11 - <<'PY'
import importlib.metadata as md
import zpe_prosody

print("zpe-prosody", md.version("zpe-prosody"))
PY

Install success only proves package acquisition/import. Product scope, stale PyPI state, platform limits, and blockers remain in the front-door sections below.

• PyPI copy is stale; retrieval and transfer blockers below are not altered by install success.

Quick Start

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
make repo-sanity
make package-sanity
make test

Optional API wrapper dependency:

python -m pip install ".[api]"

The base wheel ships only src/zpe_prosody. No CLI or historical gate harness is packaged as a runtime contract. Read docs/LEGAL_BOUNDARIES.md before widening any claim from this repo state.

---