mirror Qwen3-ForcedAligner-0.6B-GGUF from ModelScope（逐字节一致）+ 中英双语模型卡

Files mirrored byte-for-byte from https://modelscope.cn/models/42ailab/Qwen3-ForcedAligner-0.6B-GGUF (sha256 verified in-stream).

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+aligner-mmproj-f16.gguf filter=lfs diff=lfs merge=lfs -text
+aligner-backbone-f16.gguf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,126 @@
+---
+license: apache-2.0
+base_model:
+  - Qwen/Qwen3-ForcedAligner-0.6B
+base_model_relation: quantized
+pipeline_tag: automatic-speech-recognition
+library_name: gguf
+language:
+  - zh
+  - en
+tags:
+  - forced-alignment
+  - word-timestamps
+  - asr
+  - qwen3
+  - gguf
+  - 42model
+---
+
+<!--
+  README for the Hugging Face repo `42ailab/Qwen3-ForcedAligner-0.6B-GGUF`, uploaded with the package by
+  `scripts/publish-forcedaligner.sh`. Editing this file changes the repo landing page.
+-->
+# Qwen3-ForcedAligner-0.6B · Fully Local Word-Level Timestamps
+
+[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://www.apache.org/licenses/LICENSE-2.0)
+[![Runs Local](https://img.shields.io/badge/Runs-100%25_Local-brightgreen.svg)](https://42model.com)
+[![Platform](https://img.shields.io/badge/Platform-macOS_|_Windows_|_Linux-lightgrey.svg)](https://42model.com)
+[![ModelScope](https://img.shields.io/badge/ModelScope-42ailab-624AFF.svg)](https://modelscope.cn/models/42ailab/Qwen3-ForcedAligner-0.6B-GGUF)
+
+**English** | [简体中文](./README_zh.md) · Also on [ModelScope](https://modelscope.cn/models/42ailab/Qwen3-ForcedAligner-0.6B-GGUF) · [42ailab.com](https://42ailab.com)
+
+Give it an audio clip and its transcript, and this model marks **which second every character or word was spoken at** (to 80-millisecond precision) — entirely on your own computer. **No cloud, free, private.**
+
+> **The model was built by Alibaba's Qwen team and released under Apache-2.0** (Qwen3-ForcedAligner, part of the Qwen3-ASR family). **This repo is not a new model** — we converted it into a GGUF build that **runs offline on your computer**. It is an **enhancement model** and does not transcribe on its own: transcription is still done by [Qwen3-ASR](https://huggingface.co/ggml-org/Qwen3-ASR-0.6B-GGUF), and this model adds precise word-level timing on top.
+
+## 1. The problem it solves
+
+Speech recognition (ASR) tells you *what* was said, but usually not **which second each word was said at**. Many use cases depend precisely on that timing:
+
+- **Generating subtitles**: SRT / VTT subtitles need each line aligned to exact start and end times;
+- **Audio/video editing and search**: jump to the audio position for a piece of text, or cut a specific sentence precisely;
+- **Accessibility and learning**: word-by-word highlighting for read-along and listen-and-read study.
+
+Getting accurate word-level timing used to depend on a cloud service — uploading your audio to someone else's server, with the privacy concerns and possible costs that entails. This model keeps that step **entirely local**.
+
+## 2. How it works
+
+- **Input** "audio + transcript", **output** the start and end time of every character or word (at 80-millisecond frame granularity).
+- **It carries its own audio encoder**, so it **does not care which ASR you use** — download it once and pair it with any Qwen3-ASR (0.6B / 1.7B alike), instead of storing separate alignment weights per size.
+- **A single forward pass suffices** (non-autoregressive), which makes it fast; it runs locally across macOS / Windows / Linux.
+
+## 3. How well it performs
+
+On an Apple M3 Max, using a Chinese multi-speaker meeting recording (55 seconds / ~150 words / 300 word-level time points), compared point by point against the ground truth from the reference implementation (PyTorch, full precision):
+
+| Metric | Result |
+|---|---|
+| Time points **exactly matching** the reference | **84.3%** |
+| Within **±1 frame (80 ms)** of the reference | **98.7%** |
+| Tokenization **exactly aligned** with the reference | **100%** |
+| Processing speed (RTF) | **0.0063** (55 s of audio in about 300 ms) |
+
+The remaining differences come from **numerical differences** between local half precision and the reference's full precision, not from alignment errors — 84% exact matches rules out systematic drift, 98.7% within one frame rules out path errors, and 100% tokenization alignment rules out misplaced boundaries.
+
+## 4. Limitations and what's next
+
+- **Requires a transcription model**: this model only adds timestamps and does not transcribe; use it together with Qwen3-ASR.
+- **Language support**: Chinese and space-delimited languages (English and others) are verified; **Japanese / Korean** need dedicated tokenizers, so for now the model **degrades automatically and emits no word-level timestamps for them** (we would rather withhold them than give you wrong ones); support will come in a later version.
+- Next: broaden language coverage and keep improving numerical consistency.
+
+## 5. How to download and use it
+
+This model is packaged for [42model](https://42model.com), which is the recommended way to get it:
+
+**Desktop app**
+1. Download **Qwen3-ForcedAligner** from the **Model Library**;
+2. In the parameter settings of whichever Qwen3-ASR transcription model you use, turn on **Word-level timestamps**.
+
+Transcription will then emit start and end times for every character or word.
+
+## Files and license
+
+| File | Role |
+|---|---|
+| `aligner-backbone-f16.gguf` | Alignment backbone |
+| `aligner-mmproj-f16.gguf` | Audio encoder |
+| `aligner-head.bin` | Word-level classification head |
+
+Per-file sha256 values are in the bundled `manifest.json` and on the Files page, and can be verified independently.
+
+**License**: the model itself is Qwen3-ForcedAligner-0.6B, © Alibaba Qwen team, **Apache-2.0** (official sources: [Hugging Face](https://huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B) · [ModelScope](https://modelscope.cn/models/Qwen/Qwen3-ForcedAligner-0.6B) · [technical report](https://arxiv.org/abs/2601.21337)). This repo is a GGUF build converted from it and is likewise governed by [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0). By using it you agree to the upstream license terms.
+
+## Citation
+
+For **the model itself**, please cite the upstream Alibaba Qwen team (the Qwen3-ASR technical report):
+
+```bibtex
+@article{Qwen3-ASR,
+  title   = {Qwen3-ASR Technical Report},
+  author  = {Xian Shi and Xiong Wang and Zhifang Guo and Yongqi Wang and Pei Zhang and Xinyu Zhang and Zishan Guo and Hongkun Hao and Yu Xi and Baosong Yang and Jin Xu and Jingren Zhou and Junyang Lin},
+  journal = {arXiv preprint arXiv:2601.21337},
+  year    = {2026}
+}
+```
+
+If **this repo's GGUF build** was useful to you, you may additionally cite:
+
+```bibtex
+@misc{yang2026qwen3forcedalignergguf,
+  title        = {Qwen3-ForcedAligner-0.6B-GGUF: A Local Build of Qwen's Qwen3-ForcedAligner},
+  author       = {Yang, Zhiping},
+  year         = {2026},
+  howpublished = {\url{https://huggingface.co/42ailab/Qwen3-ForcedAligner-0.6B-GGUF}},
+  organization = {42ailab},
+  note         = {GGUF conversion and local deployment packaging; the model itself is Alibaba Qwen team's Qwen3-ForcedAligner (Apache-2.0). Contact: contact@42ailab.com}
+}
+```
+
+Contact us: **contact@42ailab.com**
+
+## About us
+
+**[42ailab](https://42ailab.com)** — an AI research lab exploring the boundaries of intelligence. Grounded in cognitive science, we work toward a deep integration of AI and human intelligence — to truly understand and augment intelligence, carbon-based and silicon-based alike.
+
+**[42model](https://42model.com)** — a high-performance local inference engine from 42ailab that runs translation, transcription, recognition, chat and coding on your own machine, free and private; with optional cloud compute for fine-tuning your own models and bringing them back to run locally.

README_zh.md

@@ -0,0 +1,104 @@
+<!--
+  这是 ModelScope 仓 `42ailab/Qwen3-ForcedAligner-0.6B-GGUF` 的 README，由
+  `scripts/publish-forcedaligner.sh` 随包上传。改这里 = 改仓首页。
+-->
+# Qwen3-ForcedAligner-0.6B · 全本地字级时间戳
+
+[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://www.apache.org/licenses/LICENSE-2.0)
+[![Runs Local](https://img.shields.io/badge/Runs-100%25_Local-brightgreen.svg)](https://42model.com)
+[![Platform](https://img.shields.io/badge/Platform-macOS_|_Windows_|_Linux-lightgrey.svg)](https://42model.com)
+
+给一段音频和它的文字转写，本模型标出**每个字 / 词是在第几秒说的**（精确到 80 毫秒）——全部在你自己的电脑上完成，**不上云、免费、私密**。
+
+> **模型由阿里 Qwen 团队研发、以 Apache-2.0 开源**（Qwen3-ForcedAligner，Qwen3-ASR 家族）。**本仓不是新模型**，而是我们把它转换成**能在你电脑上离线运行**的 GGUF 部署档。它是一个**增强模型**、本身不做转写：转写仍由 [Qwen3-ASR](https://modelscope.cn/models/ggml-org/Qwen3-ASR-0.6B-GGUF) 完成，本模型为其补上精确的字级时间。
+
+## 一、它解决什么问题
+
+语音转文字（ASR）能告诉你「说了什么」，却通常不告诉你「**每个字是在第几秒说的**」。而很多场景恰恰离不开这份时间信息：
+
+- **生成字幕**：SRT / VTT 字幕需要每句话对齐到精确的起止时间；
+- **音视频剪辑与检索**：按文字跳到音频对应位置、精准剪掉某一句；
+- **可访问性与学习**：逐字高亮跟读、听读对照。
+
+过去要拿到精确的字级时间，往往得依赖云端服务——把你的音频上传到别人的服务器，既有隐私顾虑，也可能产生费用。本模型让这一步**完全在本地完成**。
+
+## 二、它是怎么做到的
+
+- **输入**「音频 + 转写文本」，**输出**每个字 / 词的起止时间（80 毫秒帧粒度）。
+- **自带音频编码器**，因此**不挑你用的是哪个 ASR**——一次下载，即可搭配任意 Qwen3-ASR（0.6B / 1.7B 通用），不必为每个尺寸各存一份对齐权重。
+- **单次前向即可完成**（非自回归），速度很快；跨 macOS / Windows / Linux 本地运行。
+
+## 三、效果如何
+
+在 Apple M3 Max 上，用一段中文多人会议音频（55 秒 / 约 150 词 / 300 个字级时间点），与参考实现（PyTorch 全精度）的逐字标准答案逐一对比：
+
+| 指标 | 结果 |
+|---|---|
+| 时间点与参考**完全一致** | **84.3%** |
+| 落在参考 **±1 帧（80 毫秒）以内** | **98.7%** |
+| 分词与参考**完全对齐** | **100%** |
+| 处理速度（RTF） | **0.0063**（55 秒音频约 300 毫秒算完） |
+
+其余差异来自本地半精度与参考全精度之间的**数值差**，而非对齐机制出错——84% 完全一致排除了系统性偏移，98.7% 落在一帧内排除了路径错误，100% 分词对齐排除了对错位置。
+
+## 四、局限与下一步
+
+- **需要搭配转写模型**：本模型只补时间戳、不做转写，请配合 Qwen3-ASR 使用。
+- **语言支持**：中文、以及以空格分词的语言（英文等）已验证；**日文 / 韩文**需要专用分词器，目前会**自动降级、不输出其字级时间戳**（宁可暂不提供，也绝不给你错的），后续版本支持。
+- 下一步：扩展语言覆盖、并持续提升数值一致性。
+
+## 五、怎么下载与使用
+
+本模型专为 [活水模型（42model）](https://42model.com) 打包，推荐通过它获取：
+
+**桌面版**
+1. 在「模型库」里下载 **Qwen3-ForcedAligner**；
+2. 在你所用的 Qwen3-ASR 转录模型的「参数设置」里，打开「**字级时间戳**」开关。
+
+之后转录即输出每个字 / 词的起止时间。
+
+## 文件与许可
+
+| 文件 | 作用 |
+|---|---|
+| `aligner-backbone-f16.gguf` | 对齐主干 |
+| `aligner-mmproj-f16.gguf` | 音频编码器 |
+| `aligner-head.bin` | 字级分类头 |
+
+各文件 sha256 见随包 `manifest.json` 与 ModelScope 文件页，可自行校验。
+
+**许可**：模型本体为 Qwen3-ForcedAligner-0.6B，© 阿里 Qwen 团队，**Apache-2.0**（官方源：[ModelScope](https://modelscope.cn/models/Qwen/Qwen3-ForcedAligner-0.6B) · [技术报告](https://arxiv.org/abs/2601.21337)）。本仓为其转换的 GGUF 部署档，同样遵循 [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0)，使用即表示同意上游许可条款。
+
+## 引用
+
+**模型本体**请引用上游阿里 Qwen 团队（Qwen3-ASR 技术报告）：
+
+```bibtex
+@article{Qwen3-ASR,
+  title   = {Qwen3-ASR Technical Report},
+  author  = {Xian Shi and Xiong Wang and Zhifang Guo and Yongqi Wang and Pei Zhang and Xinyu Zhang and Zishan Guo and Hongkun Hao and Yu Xi and Baosong Yang and Jin Xu and Jingren Zhou and Junyang Lin},
+  journal = {arXiv preprint arXiv:2601.21337},
+  year    = {2026}
+}
+```
+
+若**本仓的 GGUF 量化部署档**对你有帮助，可另附���
+
+```bibtex
+@misc{yang2026qwen3forcedalignergguf,
+  title        = {Qwen3-ForcedAligner-0.6B-GGUF: A Local Build of Qwen's Qwen3-ForcedAligner},
+  author       = {Yang, Zhiping},
+  year         = {2026},
+  howpublished = {\url{https://modelscope.cn/models/42ailab/Qwen3-ForcedAligner-0.6B-GGUF}},
+  organization = {42ailab},
+  note         = {GGUF 转换与本地部署封装；模型本体为阿里 Qwen 团队 Qwen3-ForcedAligner（Apache-2.0）。Contact: contact@42ailab.com}
+}
+```
+
+联系我们：**contact@42ailab.com**
+
+## 关于我们
+
+**[活水 AI 实验室（42ailab）](https://42ailab.com)** — 探索智能边界的 AI 创新实验室，以认知科学为基石，推动 AI 与人类智能的深度融合，真正理解并增强智能 —— 碳基的，也是硅基的。
+
+**[活水模型（42model）](https://42model.com)** — 由活水 AI 实验室出品的高性能本地 AI 大模型推理引擎，让翻译、转写、识别、对话、编程等 AI 能力在你的本机免费私密运行；并可借云端算力微调专属模型，回传本机运行。

aligner-backbone-f16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ceff10e4b524bf0d683bce032e035aaaeaaa265537b55d4e2e1867601638cefd
+size 1198443296

aligner-head.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a3f5ed737b36838f078e585b8844d24a1492cf18a8a32c51092f384fc706d28
+size 20480012

aligner-mmproj-f16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73559b9e10132bdc97e001d5d383e6e72ebaf043fda4c4fab45c82ddbe0f2d67
+size 639672736

configuration.json

@@ -0,0 +1 @@
+{"framework": "pytorch", "task": "text-generation", "allow_remote": true}

manifest.json

@@ -0,0 +1,9 @@
+{
+  "repo": "42ailab/Qwen3-ForcedAligner-0.6B-GGUF",
+  "files": [
+    {"path": "README.md", "sha256": "978ee19f7d2aea491aae6f92db83b7e2ae7ddeec59a919650c379fb73f4343ad", "size": 5437},
+    {"path": "aligner-backbone-f16.gguf", "sha256": "ceff10e4b524bf0d683bce032e035aaaeaaa265537b55d4e2e1867601638cefd", "size": 1198443296},
+    {"path": "aligner-head.bin", "sha256": "8a3f5ed737b36838f078e585b8844d24a1492cf18a8a32c51092f384fc706d28", "size": 20480012},
+    {"path": "aligner-mmproj-f16.gguf", "sha256": "73559b9e10132bdc97e001d5d383e6e72ebaf043fda4c4fab45c82ddbe0f2d67", "size": 639672736}
+  ]
+}