humyn-labs/Indic-High-Fidelity-MultiSpeaker-ASR
收藏Hugging Face2026-03-14 更新2026-04-05 收录
下载链接:
https://hf-mirror.com/datasets/humyn-labs/Indic-High-Fidelity-MultiSpeaker-ASR
下载链接
链接失效反馈官方服务:
资源简介:
---
license: cc-by-4.0
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
dataset_info:
features:
- name: language
dtype: string
- name: file_name
dtype: string
- name: audio
dtype: audio
- name: transcript_json
dtype: string
splits:
- name: train
num_bytes: 1065664379
num_examples: 37
download_size: 1064810251
dataset_size: 1065664379
task_categories:
- automatic-speech-recognition
language:
- hi
- ml
- mr
- te
- ta
- bn
- kn
- bh
- as
- gu
- pa
tags:
- ASR
- Conversational-speech
- multi-speaker
- indic-languages
size_categories:
- n<1K
---
# Dataset Overview
This dataset contains high-quality multi-speaker conversational audio recordings curated for Automatic Speech Recognition (ASR) research across multiple Indic languages.
The dataset includes:
- Paired audio + timestamped transcripts
- Natural, non-scripted conversational speech
- Dual-speaker interactions
- Segment-level speaker annotations
- Regionally diverse accents
# Audio Specifications
- Format: WAV (PCM 16-bit)
- Sampling Rate: 16 kHz
- Channel: Mono
- Speech Type: Natural conversational dialogue
- Recording Style: Dual-speaker spontaneous interaction
- Typical Duration: 10–30 minutes per recording
All audio files are normalized to ensure consistent duration reporting and playback compatibility.
# Supported Languages
This dataset includes conversational speech recordings in:
- Assamese
- Odia
- Bengali
- Bhojpuri
- Chhattisgarhi
- Gujarati
- Haryanvi
- Hindi
- Punjabi
- Marathi
- Tamil
- Kannada
- Malayalam
- Telugu
The dataset preserves natural accent variation and conversational speech characteristics.
# Speaker Representation
- Dual-speaker conversational recordings
- Natural, spontaneous dialogue
- Regionally representative speakers
- Conversational turn-taking preserved
# Dataset Creation Methodology
## Data Collection
Speech data was collected from native speakers across multiple Indian regions to ensure:
- Accent diversity
- Natural conversational flow
- Real-world dialogue patterns
- Informal and semi-formal speech contexts
Topics include:
- Everyday life discussions
- Social interactions
- Business and finance
- Public affairs
- General conversational topics
# Transcription Process
- Manual transcription by native speakers
- Reviewed for linguistic accuracy
- Timestamp-level segmentation
- Speaker-labeled segments
- Preserves conversational fillers and natural pauses
Each transcript entry contains:
- start timestamp
- end timestamp
- speaker label
- text content
# Intended Use
Designed for:
- Training and fine-tuning ASR models
- Conversational ASR benchmarking
- Speaker diarization research
- Speaker turn detection
- Multi-speaker modeling
- Academic and open research
# Out-of-Scope Uses
This dataset is not intended for:
- Safety-critical or real-time production systems without additional validation
- Commercial deployment without attribution (CC BY 4.0 required)
- Medical, clinical, legal, or diagnostic applications
# License
Creative Commons Attribution 4.0 International (CC BY 4.0)
📬 Contact
For dataset-related queries, please contact:-
[support@humynlabs.ai]
license: CC BY 4.0(知识共享署名4.0国际许可协议)
configs:
- config_name: 默认(default)
data_files:
- split: 训练集(train)
path: data/train-*
dataset_info:
features:
- name: 语言(language)
dtype: 字符串(string)
- name: 文件名(file_name)
dtype: 字符串(string)
- name: 音频(audio)
dtype: 音频(audio)
- name: 转录JSON(transcript_json)
dtype: 字符串(string)
splits:
- name: 训练集(train)
num_bytes: 1065664379
num_examples: 37
download_size: 1064810251
dataset_size: 1065664379
task_categories:
- 自动语音识别(automatic-speech-recognition)
language:
- 印地语(hi)
- 马拉雅拉姆语(ml)
- 马拉地语(mr)
- 泰卢固语(te)
- 泰米尔语(ta)
- 孟加拉语(bn)
- 卡纳达语(kn)
- 博杰普尔语(bh)
- 阿萨姆语(as)
- 古吉拉特语(gu)
- 旁遮普语(pa)
tags:
- ASR(自动语音识别,Automatic Speech Recognition)
- 会话语音(Conversational-speech)
- 多说话人(multi-speaker)
- 印度语系(indic-languages)
size_categories:
- 样本数少于1000(n<1K)
# 数据集概览
本数据集包含高质量多说话人会话音频录制数据,专为多印度语系下的自动语音识别(Automatic Speech Recognition, ASR)研究打造。
本数据集包含:
- 配对音频与带时间戳的转录文本
- 自然、非脚本化的会话语音
- 双说话人交互场景
- 片段级说话人标注
- 区域多样化口音
# 音频规格
- 格式:WAV(PCM 16位)
- 采样率:16 kHz
- 声道:单声道
- 语音类型:自然会话对话
- 录制方式:双说话人自发交互
- 典型时长:每份录制10–30分钟
所有音频文件均经过归一化处理,以确保时长报告统一及播放兼容性。
# 支持语言
本数据集包含以下语言的会话语音录制数据:
- 阿萨姆语
- 奥里亚语
- 孟加拉语
- 博杰普尔语
- 恰蒂斯加尔语
- 古吉拉特语
- 哈里亚纳维语
- 印地语
- 旁遮普语
- 马拉地语
- 泰米尔语
- 卡纳达语
- 马拉雅拉姆语
- 泰卢固语
本数据集保留了自然的口音变化与会话语音特征。
# 说话人表征
- 双说话人会话录制
- 自然自发的对话
- 具有区域代表性的说话人
- 完整保留会话轮次转换
# 数据集构建方法
## 数据采集
语音数据采集自印度多个地区的母语使用者,以确保:
- 口音多样性
- 自然会话流畅性
- 真实对话模式
- 非正式与半正式的语音场景
话题涵盖:
- 日常生活讨论
- 社交互动
- 商业与金融
- 公共事务
- 通用会话话题
# 转录流程
- 由母语使用者进行人工转录
- 经语言准确性审核
- 时间戳级别的分段标注
- 带说话人标签的语音片段
- 保留会话填充词与自然停顿
每条转录条目包含:
- 起始时间戳
- 结束时间戳
- 说话人标签
- 文本内容
# 预期用途
本数据集专为以下场景设计:
- 自动语音识别模型的训练与微调
- 会话式自动语音识别基准测试
- 说话人 diarization(Speaker Diarization)研究
- 说话人轮次检测
- 多说话人建模
- 学术与开放研究
# 非适用场景
本数据集不适用于:
- 未经额外验证的安全关键型或实时生产系统
- 未经署名的商业部署(需遵守CC BY 4.0许可协议)
- 医疗、临床、法律或诊断类应用
# 许可协议
知识共享署名4.0国际许可协议(Creative Commons Attribution 4.0 International, CC BY 4.0)
📬 联系方式
若有数据集相关疑问,请联系:
[support@humynlabs.ai]
提供机构:
humyn-labs


