Dataset

Dataset Name	Type	#Samples	# Hours	Task	Metrics
MDRM-test [1]	Short Clips	22,208	87	short financial clip ASR	WER
SPGISpeech-test [2]	Short Clips	39,341	130	short financial clip ASR	WER
Earning-21 [3]	Long Audio	44	39	long financial audio ASR	WER
Earning-22 [4]	Long Audio	125	120	long financial audio ASR	WER
FinAudioSum	Long Audio	64	55	long financial audio Summarization	Rouge-L & BERTScore

We create FinAudioSum based on the ECTSum dataset, originally designed for earnings call summarization using textual data. ECTSum comprises 2,425 earnings transcripts paired with expert-generated, telegram-style summaries. We obtain corresponding audio recordings for the ECTSum test set from earningscast. Overlapping recordings with Earnings-21 and Earnings-22 (spanning 2019–2022) are removed. The final FinAudioSum dataset includes 64 recordings totaling 55 hours.

[1] [What You Say and How You Say It Matters: Predicting Financial Risk Using Verbal and Vocal Cues](https://aclanthology.org/P19-1038.pdf) ACL 2019. [Data](https://github.com/GeminiLn/EarningsCall_Dataset/tree/master)

[2] [SPGISpeech: A Large-Scale Dataset for Financial Speech Recognition](https://arxiv.org/pdf/2104.02014) InterSpeech 2021.

[3] [Earnings-21: A practical benchmark for ASR in the wild](https://arxiv.org/pdf/2104.11348) InterSpeech 2021.

[4] Earnings-22: A Practical Benchmark for Accents in the Wild