tintitu/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404-onnx

Paraformer-Large-Contextual 热词增强语音识别模型 (ONNX 量化版)

面向中文自动语音识别（ASR）的端到端非自回归大模型（集成上下文热词偏置网络 Contextual Biasing Network），当前目录包含 ONNX 量化主模型、热词偏置嵌入网络、分词词典、分词词表及解码配置。

模型信息

• —模型名称：Paraformer-Large-Contextual ASR (NAT-ZH-CN-16k-Common-Vocab8404)
• —适用任务：离线语音识别（ASR，支持动态热词/专有名词增强）
• —模型架构：非自回归 Transformer + Contextual Biasing Network
• —支持语言：中文普通话（zh-cn）
• —词表规模：8404 常用汉字与特殊 Token
• —特色能力：支持在推理时动态注入热词列表（如人名、地名、行业术语、自定义品牌词），显著提升特定词汇的召回率与准确率。
• —主要文件：
• —model_quant.onnx：经过 INT8 静态量化的声学主干推理模型
• —model_eb.onnx：热词嵌入偏置网络（Embedding Biaser）
• —seg_dict：热词切词与分词辅助字典
• —tokens.json：字符与词元 Token 映射词表
• —config.yaml：模型推理与热词协同解码配置
• —am.mvn：声学特征 CMVN（倒谱均值方差归一化）统计文件

硬件资源门槛与性能预期 (CPU)

• —内存常驻与峰值：约 1.1 GB ~ 1.6 GB RAM（包含热词偏置网络与分词字典常驻，推理峰值约 1.4 GB）
• —性能预期 (RTF)：典型 8 核 CPU 上，无热词或百词以内小热词表时 RTF 预期约为 0.12 ~ 0.22（处理 10 秒音频耗时约 1.2 ~ 2.2 秒）
• —推荐配置：4 核及以上 CPU、8 GB 及以上系统物理内存

音频输入规范

• —采样率要求：16000 Hz（16 kHz）
• —声道配置：单声道（Mono）
• —音频格式：16-bit PCM 或归一化浮点格式（数值范围 [-1.0, 1.0]）
• —特征提取：默认提取 80 维 Fbank 特征，配合 CMVN 归一化输入

推荐流水线协作

• —前置模块：推荐搭配 speech_fsmn_vad_zh-cn-16k-common-onnx 进行语音活动检测与长音频断句，滤除环境静音并提高识别效率。
• —后置模块：识别输出为纯文字流，推荐后置串联标点恢复模型（如 punc_ct-transformer_zh-cn-common-vocab272727-onnx）完成自动标点补充。
• —热词调用：热词输入通常以空格或换行分隔；若不传入热词，模型表现与标准版 Paraformer-Large 基本一致。

上游来源

• —上游开源项目：`alibaba-damo-academy/FunASR`
• —ModelScope 官方仓库：`damo/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404-onnx`
• —推理引擎兼容：ONNX Runtime、sherpa-onnx、FunASR Runtime

本目录包含的是上游模型导出并量化后的 ONNX 推理版本，版权归原作者及阿里巴巴达摩院/通义实验室所有。

许可证与再分发

遵循上游 FunASR 与 ModelScope 开源协议（Apache-2.0 / ModelScope 社区许可）。商业集成或二次分发时，请保留上游版权声明、许可说明及原始出处链接。

文件校验清单 (SHA-256)

下载后请核对文件大小和 SHA256 校验和。

Paraformer-Large-Contextual ASR Model (ONNX Quantized) (English Documentation)

An end-to-end non-autoregressive speech recognition model with contextual hotword biasing for Mandarin Chinese ASR. This directory contains quantized ONNX weights, embedding biaser network, segmentation dictionary, vocabulary mappings, and configuration files.

Model Information

• —Model Name: Paraformer-Large-Contextual ASR (NAT-ZH-CN-16k-Common-Vocab8404)
• —Task: Offline Automatic Speech Recognition (ASR) with dynamic hotword enhancement
• —Architecture: Non-Autoregressive Transformer + Contextual Biasing Network
• —Language: Mandarin Chinese (zh-cn)
• —Vocabulary Size: 8,404 Chinese characters and special tokens
• —Key Feature: Supports runtime hotword injection (e.g., entity names, specialized terms, branded keywords) to substantially improve keyword recall and precision.
• —Primary Files:
• —model_quant.onnx: INT8 statically quantized acoustic inference backbone
• —model_eb.onnx: Embedding biaser network for contextual hotword injection
• —seg_dict: Word segmentation dictionary for hotword tokenization
• —tokens.json: Character and token mapping dictionary
• —config.yaml: Model configuration and biaser decoding parameters
• —am.mvn: Cepstral mean and variance normalization (CMVN) statistics

Hardware Footprint & Sizing (CPU)

• —Memory Footprint (RAM): ~1.1 GB to 1.6 GB (Including biaser model and dictionary cache, peak ~1.4 GB)
• —Expected RTF (Real-Time Factor): RTF of ~0.12 - 0.22 on typical 8-core CPUs with standard hotword lists (processing a 10s audio clip takes ~1.2s - 2.2s)
• —Recommended System: 4+ CPU cores, 8+ GB RAM

Audio Input Specifications

• —Sampling Rate: 16000 Hz (16 kHz)
• —Channels: Single channel (Mono)
• —Format: 16-bit PCM or normalized float array within [-1.0, 1.0]
• —Feature Extraction: 80-dimensional log Mel-filterbank features with CMVN

Recommended Pipeline Integration

• —Upstream Preprocessing: Recommended pairing with speech_fsmn_vad_zh-cn-16k-common-onnx for voice activity detection and sentence segmentation.
• —Downstream Postprocessing: Paraformer outputs raw text without punctuation. It is recommended to cascade a punctuation restoration model (such as punc_ct-transformer_zh-cn-common-vocab272727-onnx).
• —Hotword Usage: Hotword inputs are typically supplied as whitespace- or newline-separated lists. When no hotwords are provided, performance is equivalent to standard Paraformer-Large.

Upstream Sources

• —Upstream Project: `alibaba-damo-academy/FunASR`
• —ModelScope Repository: `damo/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404-onnx`
• —Supported Inference Engines: ONNX Runtime, sherpa-onnx, FunASR Runtime

This directory provides a quantized ONNX inference distribution derived from upstream checkpoints. All copyrights belong to the original authors and Alibaba DAMO Academy / Tongyi Lab.

Licensing and Redistribution

Governed by the upstream FunASR and ModelScope licensing terms (Apache-2.0 / ModelScope Community License). Please retain upstream copyright statements, notices, and source links in downstream redistributions.

File Verification Manifest (SHA-256)

Verify file sizes and SHA-256 checksums after downloading.