backpack-run/Kokoro-82M-Backpack-TTS

Kokoro 82M — Backpack Voice Package

🎒 Backpack Verified

Lightweight local text-to-speech with a curated set of voices. This package stages immutable upstream artifacts for Backpack's voice runtime layer. It does not replace the chat model selected by the user.

Package

Field	Value
Capability	text-to-speech
Input	text
Output	audio
Runtime	kokoro
Configured runtime revision	kokoro==0.9.4;en_core_web_sm==3.8.0

| Format | pytorch | | Precision | F32 | | Package size | 314.1 MiB | | Recommended RAM | 1.88 GB | | Languages | multilingual |

Validation status

The packager verified the immutable revision, selected-file inventory, non-empty files, hashes, configuration JSON, and the primary artifact container/header. It also loaded the package and passed deterministic audio inference with the configured runtime.

Integrity	Metadata	Runtime load	Audio inference	Tokenizer
passed	passed	passed	passed	passed

Run with Kokoro

Install kokoro==0.9.4, construct KModel from config.json and kokoro-v1_0.pth, then pass one of the packaged voices/*.pt files to KPipeline.

Provenance

• Upstream: hexgrad/Kokoro-82M

• Immutable revision: f3ff3571791e39611d31c381e3a41a3af07b4987

• License: apache-2.0

• Backpack copied the selected upstream artifacts without modifying model weights.

• Backpack did not train this model and does not claim ownership of it.

Files and checksums

• config.json — 2.3 KiB — 5abb01e2403b072bf03d04fde160443e209d7a0dad49a423be15196b9b43c17f
• kokoro-v1_0.pth — 312.1 MiB — 496dba118d1a58f5f3db2efc88dbdc216e0483fc89fe6e47ee1f2c53f18ad1e4
• voices/af_heart.pt — 511.2 KiB — 0ab5709b8ffab19bfd849cd11d98f75b60af7733253ad0d67b12382a102cb4ff
• voices/am_michael.pt — 511.2 KiB — 9a443b79a4b22489a5b0ab7c651a0bcd1a30bef675c28333f06971abbd47bd37
• voices/bf_emma.pt — 511.2 KiB — d0a423deabf4a52b4f49318c51742c54e21bb89bbbe9a12141e7758ddb5da701
• voices/bm_george.pt — 511.2 KiB — f1bc812213dc59774769e5c80004b13eeb79bd78130b11b2d7f934542dab811b

Review the upstream model card and license before use or redistribution. Speech systems can mis-transcribe, synthesize misleading content, or behave differently across languages and accents.