LibriSpeech
A large-scale English speech recognition benchmark dataset derived from LibriVox audiobooks. It contains approximately 1,000 hours of speech data and is one of the most widely used standard evaluation sets in the automatic speech recognition field.
Timeline (last 90 days)
On the LibriSpeech test-clean/test-other datasets, Oído's on-device computation achieves a word error rate of 3.70% / 8.23%, differing from the full-precision version by only 0.12 percentage points.
LibriSpeech consists primarily of read speech, while MSP-Podcast is emotionally rich podcast conversation, and the two differ markedly in style.
The routing organizational structure learned on LibriSpeech can be directly transferred to the MSP-Podcast dataset without retraining at the representation level.
The experimental data used in the study includes Arabic native and non-native children's speech, the English MyST children's speech dataset, and the MGB-2 and LibriSpeech test-clean adult benchmark data.
On the LibriSpeech benchmark, Oído's word error rate (WER) is 3.7 (clean) and 8.2 (other), while Whisper tiny.en on a laptop achieves 6.3 and 15.9.
PTC-Bias current experiments are primarily based on the English LibriSpeech dataset and specific SpeechLLMs; its performance in other languages such as Chinese, real noisy environments, and larger bias word lists remains to be validated.
The paper validates on the LibriSpeech dataset, covering two different SpeechLLMs, with the maximum bias word list size reaching 2,000 words.
Conformer achieved approximately 15% relative error rate reduction on the LibriSpeech benchmark compared to a pure Transformer architecture.
Current state-of-the-art systems have achieved a WER below 5% on the standard English test set LibriSpeech.
在 LibriSpeech 基准测试上顶尖系统 WER 已低于 2%，在 Switchboard 对话场景中通常为 5-8%，在多人会议场景（AMI、ICSI）中可能高达 15-25%
All Facts (11)
Current state-of-the-art systems have achieved a WER below 5% on the standard English test set LibriSpeech.
70%UnverifiedOn the LibriSpeech test-clean/test-other datasets, Oído's on-device computation achieves a word error rate of 3.70% / 8.23%, differing from the full-precision version by only 0.12 percentage points.
50%UnverifiedLibriSpeech consists primarily of read speech, while MSP-Podcast is emotionally rich podcast conversation, and the two differ markedly in style.
50%UnverifiedThe routing organizational structure learned on LibriSpeech can be directly transferred to the MSP-Podcast dataset without retraining at the representation level.
50%UnverifiedThe experimental data used in the study includes Arabic native and non-native children's speech, the English MyST children's speech dataset, and the MGB-2 and LibriSpeech test-clean adult benchmark data.
50%UnverifiedOn the LibriSpeech benchmark, Oído's word error rate (WER) is 3.7 (clean) and 8.2 (other), while Whisper tiny.en on a laptop achieves 6.3 and 15.9.
50%UnverifiedPTC-Bias current experiments are primarily based on the English LibriSpeech dataset and specific SpeechLLMs; its performance in other languages such as Chinese, real noisy environments, and larger bias word lists remains to be validated.
50%UnverifiedThe paper validates on the LibriSpeech dataset, covering two different SpeechLLMs, with the maximum bias word list size reaching 2,000 words.
50%UnverifiedConformer achieved approximately 15% relative error rate reduction on the LibriSpeech benchmark compared to a pure Transformer architecture.
50%Unverified在 LibriSpeech 基准测试上顶尖系统 WER 已低于 2%，在 Switchboard 对话场景中通常为 5-8%，在多人会议场景（AMI、ICSI）中可能高达 15-25%
50%UnverifiedLibriSpeech、CommonVoice 等主流语音识别训练集大多来自朗读录音或受控环境，信噪比高、语速均匀、句子完整
50%