xumo/reader-speech-models

Reader accepted local speech artifacts

Reader-converted artifacts, not an official Qwen release. Files retain the bytes accepted by Reader; models.lock.json pins size and SHA-256. Bind downloads to the full publication commit, never a moving main branch.

Included artifacts

• asr-0.6b-f16
• asr-1.7b-f16
• asr-1.7b-q4-k-m
• asr-1.7b-q8-0
• aligner-q8-0
• mmproj-asr-0.6b-reader-compat
• mmproj-asr-1.7b-reader-compat
• mmproj-forcedaligner-reader-compat
• qwen3-tts-0.6b-base-f16
• synthetic-heart
• synthetic-george

Sources and conversion

Qwen Base revision: 5d83992436eae1d760afd27aff78a71d676296fc. Converter: ggml-org/llama.cpp 67107ac34f62882d8b53ea41ad68010ea635b9be. Reader packaging source commit: 14d2029ba099c97ea4b2544bef0c58de4f11ed79 (may precede its source-repository push). The complete recipe and reconstruction evidence are included in this model repo. Pinned public converter: https://github.com/ggml-org/llama.cpp/blob/67107ac34f62882d8b53ea41ad68010ea635b9be/convert_hf_to_gguf.py

The Qwen TTS GGUF pair includes the talker and projector/codec. The generated reader-model.json preserves the original source-file receipt; raw safetensors listed in that receipt are source inputs and are not required for inference. Rebuilding is a separate operation; consult conversion.lock.json and rebuild.json. Use the locked Python/package versions, install the converter's pinned gguf-py, download all raw files from the pinned Base revision, and preserve the source directory basename recorded by the recipe (it can affect GGUF naming metadata). Run the two argument sets in the recipe through the pinned converter, providing the raw model directory as positional input and fresh output paths. Keep the speech_tokenizer subdirectory beside the raw model. Reconstruction is CPU-only; the accepted record uses OMP/MKL/OpenBLAS thread counts of 4 and offline HF mode.

Synthetic Heart/George WAVs, when included, are reference audio resources, not model weights or user recordings. Their Kokoro model revision and source voice are recorded in each artifact origin. Preserve the supplied WAV bytes to retain the accepted speaker conditioning; generating new reference WAVs creates new assets.

Runtime compatibility and limits

Use the Reader runtime matching the locked upstream revision. Reader validates Qwen synthesis for English/Chinese; other upstream languages are not claimed by this release. Validated host: Linux x64, GTX 1080 Ti, CUDA SM61. Model files and platform runtime binaries have separate lifecycles. This repository contains no runtime binary, personal recordings, account data, or historical ASR/FA files with unverified provenance.

Attribution

Qwen3-TTS: https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base Kokoro reference source: https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX Upstream model cards mark Apache-2.0. See LICENSE. Reader changes: conversion to F16 GGUF and explicit packaging/identity metadata. The publication commit is not a model retraining or a new quality claim.

ASR/FA conversion provenance

The accepted text GGUF and Reader-compatible projectors were reconstructed exactly from the official revisions and raw-file hashes in asr-conversion.lock.json. asr-rebuild.json compares all eight outputs to their accepted SHA values.

To reconstruct: obtain public ggml-org/llama.cpp commit 66d65ec29ba7c440cbc31b6f63b74a17b536ba65, apply the included legacy-asr-conversion.patch, and check the four source hashes in legacy-asr-converter.lock.json. This freezes conversion-only sources; it does not restore an old inference backend. Use the locked Python environment and the patched source's gguf-py. Convert each official raw model to F16 text GGUF and an F16 mmproj. The original model folder basename affects metadata and must match.

Build the CPU llama-quantize tool from current upstream commit 67107ac34f62882d8b53ea41ad68010ea635b9be. Quantize ASR 1.7B F16 to Q8_0/Q4_K_M and FA F16 to Q8_0 with four threads. Apply the included standalone Reader projector converter (conversion/) with --encoder-compat, using gguf-py from that current pinned commit. No tensors are retrained; compare final SHA values against the model lock before use.

The Reader-compatible mmproj files require Reader's encoder/aligner compatibility adapter. Ordinary upstream projectors are not interchangeable with these accepted files. See the per-artifact origins for source revisions and compatibility tags. Converter code is covered separately by LICENSE.llama.cpp; model provenance and Apache-2.0 attribution remain in LICENSE and the official Qwen model cards.