FinASR-Bench

FinASR-Bench is a multilingual benchmark for evaluating and adapting automatic speech recognition (ASR) systems in the financial domain. It contains both real-world financial speech and synthetically generated financial speech in three languages:

• English (EN)
• Chinese (ZH)
• Japanese (JA)

Unlike conventional ASR benchmarks that mainly evaluate transcription accuracy using WER or CER, FinASR-Bench is designed to support evaluation of financially critical recognition errors, including errors involving numerical values, currencies, percentages, financial entities, fiscal periods, negation, comparison, and relations among financial facts.

The benchmark is designed to be used together with Structured Financial Error Rate (FER), a financial-domain ASR evaluation metric that explicitly measures errors in structured financial facts.

---

1. Dataset Overview

FinASR-Bench contains six dataset configurations:

Configuration	Type	Language	Train	Validation	Test	Total Hours
real-en	Real	English	1,153	133	124	6.20
real-zh	Real	Chinese	3,000	300	150	3.90
real-ja	Real	Japanese	3,000	300	150	11.63
synthetic-en	Synthetic	English	18,795	2,000	990	116.90
synthetic-zh	Synthetic	Chinese	18,851	2,000	1,000	115.13
synthetic-ja	Synthetic	Japanese	18,285	2,000	999	120.98

The real-world subsets are constructed from existing speech corpora, while the synthetic subsets contain controlled financial-domain utterances synthesized using TTS.

Real-world data sources

The real-world subsets are constructed from the following speech resources:

• English: SPGISpeech, SPGISpeech2, Earnings-22, and Earnings-25
• Chinese: WenetSpeech
• Japanese: ReazonSpeech

These subsets are selected and processed to provide multilingual real-world ASR evaluation in financial or financially relevant speech conditions.

Synthetic data

The synthetic subsets are generated from manually designed financial-domain text using CosyVoice3 zero-shot speech synthesis.

For speaker diversity, 20 reference speakers are used for each language:

• English reference speakers: SPGISpeech
• Chinese reference speakers: AISHELL-1
• Japanese reference speakers: ReazonSpeech

The synthetic data are specifically designed to stress ASR systems on financially important information that may not be sufficiently represented in conventional ASR benchmarks.

---

2. Loading the Dataset

FinASR-Bench is hosted as a multi-configuration Hugging Face dataset.

Install the required packages:

pip install datasets huggingface_hub

Then load a specific configuration:

from datasets import load_dataset

dataset = load_dataset(
    "wuxianchao/FinASR-Bench",
    "real-en",
)

print(dataset)

For example:

train_set = dataset["train"]
validation_set = dataset["validation"]
test_set = dataset["test"]

print(len(train_set))
print(len(validation_set))
print(len(test_set))

To load the Japanese synthetic subset:

dataset = load_dataset(
    "wuxianchao/FinASR-Bench",
    "synthetic-ja",
)

Available configurations are:

real-en
real-zh
real-ja
synthetic-en
synthetic-zh
synthetic-ja

---

3. Dataset Structure

A FinASR-Bench sample contains ASR transcription information together with metadata and structured financial annotations.

Representative fields include:

id
original_id
audio
text
language
dataset_type
source
speaker
duration_s
category
critical_tokens
financial_facts
derived_constraints
relations
severity_tags

A simplified example is:

{
    "id": "sample_000001",
    "original_id": "...",

    "audio": {
        "path": "sample_000001.wav",
        "bytes": b"..."
    },

    "text": "Bank of America reported net income of $20.7 billion.",

    "language": "en",
    "dataset_type": "synthetic",
    "source": "cosyvoice3_synthetic",
    "speaker": "...",
    "duration_s": 6.42,

    "category": "numbers_currency_percentage",

    "critical_tokens": [
        "Bank of America",
        "net income",
        "$20.7 billion"
    ],

    "financial_facts": [...],
    "derived_constraints": [...],
    "relations": [...],
    "severity_tags": [...]
}

Not every field is necessarily populated for every sample. In particular, the availability and density of structured financial annotations depend on the source and construction procedure.

---

4. Audio Representation

To improve portability during dataset construction and avoid dependence on a particular local audio decoding backend, audio is stored as:

{
    "path": "example.wav",
    "bytes": <raw WAV bytes>
}

Therefore:

sample = dataset["train"][0]

print(sample["audio"]["path"])
print(type(sample["audio"]["bytes"]))

The raw WAV file can be recovered directly:

audio = sample["audio"]

with open(audio["path"], "wb") as f:
    f.write(audio["bytes"])

For example:

sample = dataset["train"][0]

with open("/tmp/example.wav", "wb") as f:
    f.write(sample["audio"]["bytes"])

The resulting file is a standard WAV file and can subsequently be loaded using soundfile, librosa, torchaudio, or another audio library.

For example, using soundfile:

import io
import soundfile as sf

sample = dataset["train"][0]

waveform, sample_rate = sf.read(
    io.BytesIO(sample["audio"]["bytes"])
)

print(waveform.shape)
print(sample_rate)

This approach allows users to choose their preferred audio decoding stack.

---

5. Synthetic Financial Challenge Categories

The synthetic portion of FinASR-Bench is constructed around controlled financial ASR challenges.

The twelve major challenge categories are:

Category	Target Proportion	Main Challenge
numbers_currency_percentage	15%	Numerical values, currency amounts, percentages
financial_terms	10%	Specialized financial terminology
acronyms	8%	Financial and corporate acronyms
companies_tickers_entities	7%	Company names, tickers, financial entities
number_unit_term_composition	10%	Number-unit-financial-term combinations
financial_entity_relations	12%	Relations between entities and financial metrics
confusable_financial_words	5%	Acoustically or semantically confusable financial terms
dates_fiscal_periods	6%	Dates, quarters, fiscal years, reporting periods
code_switching	5%	Mixed-language financial expressions
colloquial_finance	4%	Colloquial financial expressions
negation_condition_comparison	10%	Negation, conditions, comparisons
long_context_multi_entity	8%	Long-context reasoning and multiple financial entities

The actual distribution may differ slightly from the target distribution after synthesis and filtering.

For example, a sample can be selected by category as follows:

subset = dataset["train"].filter(
    lambda x: x["category"] == "negation_condition_comparison"
)

This makes it possible to evaluate ASR systems not only globally, but also by specific financial challenge type.

---

6. Structured Financial Facts

A central feature of FinASR-Bench is the representation of financially important information as structured facts.

A financial fact can conceptually be represented as:

$$f=(e,m,v,u,c,t,n,o,d),f\; =\; (e,m,v,u,c,t,n,o,d),$$

where:

• (e): financial entity
• (m): financial metric
• (v): normalized numerical value
• (u): unit
• (c): currency
• (t): fiscal or temporal scope
• (n): negation
• (o): comparison operator
• (d): direction

For example, the sentence:

Bank of America reported net income of $20.7 billion in Q1 2027.

may contain a fact conceptually corresponding to:

{
    "entity": "Bank of America",
    "metric": "net income",
    "value": 20.7,
    "unit": "billion",
    "currency": "USD",
    "time": "Q1 2027",
    "negation": False
}

These annotations enable evaluation beyond surface transcription similarity.

---

7. Critical Tokens

The critical_tokens field identifies expressions that are especially important for preserving financial meaning.

For example:

sample["critical_tokens"]

may return:

[
    "Bank of America",
    "net income",
    "$20.7 billion",
    "Q1 2027"
]

These tokens can be used for:

• financial entity recall;
• number recognition accuracy;
• terminology analysis;
• targeted ASR error analysis;
• construction of financially weighted evaluation metrics.

---

8. Structured Financial Error Rate

FinASR-Bench is designed to support evaluation using Structured Financial Error Rate (FER).

Traditional WER/CER treats transcription errors primarily according to edit operations. However, financial ASR errors can have substantially different semantic consequences.

For example:

Reference:
Net income increased by 15%.

Hypothesis A:
Net income increased by 16%.

Hypothesis B:
Net income decreased by 15%.

Both hypotheses may contain a small number of surface-level recognition errors, but Hypothesis B changes the financial direction and therefore has a different financial interpretation.

FER addresses this problem by comparing structured financial facts.

---

8.1 Magnitude-Aware Numerical Error

Numerical differences are measured in log space:

$$E_v(v,\widehat{v})=\min (1,\frac{\mid \log (\mathrm{\mid }v\mathrm{\mid }+ϵ)-\log (\mathrm{\mid }\widehat{v}\mathrm{\mid }+ϵ)\mid }{\tau }),E\_v(v,\backslash hat\{v\})\; =\; \backslash min\backslash left(\; 1,\; \backslash frac\{\; \backslash left|\; \backslash log(|v|+\backslash epsilon)\; -\; \backslash log(|\backslash hat\{v\}|+\backslash epsilon)\; \backslash right|\; \}\{\; \backslash tau\; \}\; \backslash right),$$

where:

$$\tau =\log 10.\backslash tau\; =\; \backslash log\; 10.$$

This makes the metric sensitive to order-of-magnitude errors while keeping the numerical error bounded.

---

8.2 Binding Error

Financial values must also be associated with the correct entity and metric.

A binding error can be represented as:

$$E_b=\mathbf{1}[(e,m,v)\mathrm{\ne }(\widehat{e},\widehat{m},\widehat{v})]\mathrm{.}E\_b\; =\; \backslash mathbf\{1\}\; \backslash left[\; (e,m,v)\; \backslash neq\; (\backslash hat\{e\},\backslash hat\{m\},\backslash hat\{v\})\; \backslash right].$$

This is particularly important in utterances containing multiple companies, metrics, periods, or numerical values.

---

8.3 Severity-Aware Evaluation

FER assigns different severity weights to different types of financial errors.

The current weighting scheme emphasizes errors such as:

• incorrect numerical value;
• incorrect entity-metric-value binding;
• negation errors;
• comparison errors;
• direction errors.

Other components include:

• unit;
• currency;
• temporal scope;
• entity;
• metric.

The resulting FER is normalized to:

$$0\le \mathrm{F}\mathrm{E}\mathrm{R}\le 1,0\; \backslash leq\; \backslash mathrm\{FER\}\; \backslash leq\; 1,$$

where lower values indicate better preservation of financially relevant information.

---

9. Recommended Evaluation Metrics

We recommend reporting conventional ASR metrics together with financially oriented metrics.

For English:

WER
Normalized WER
Semantic Distance
FER

For Chinese and Japanese:

CER
Normalized CER
Semantic Distance
FER

This provides complementary views of system performance:

• WER/CER: surface transcription accuracy
• Normalized WER/CER: accuracy after text normalization
• Semantic Distance: overall semantic similarity
• FER: preservation of structured financial information

A system can therefore improve conventional transcription accuracy without necessarily improving financially critical recognition, and vice versa.

---

10. Example ASR Evaluation Workflow

A typical evaluation workflow is:

FinASR-Bench audio
        |
        v
     ASR Model
        |
        v
Hypothesis Transcript
        |
        +-----------------------+
        |                       |
        v                       v
    WER / CER            Financial Fact
                           Extraction
                                |
                                v
                         Fact Alignment
                                |
                                v
                              FER

A minimal evaluation loop can begin with:

from datasets import load_dataset

dataset = load_dataset(
    "wuxianchao/FinASR-Bench",
    "real-en",
)

for sample in dataset["test"]:
    reference = sample["text"]
    audio_bytes = sample["audio"]["bytes"]

    # 1. Decode audio
    # 2. Run your ASR model
    # hypothesis = asr(...)

    # 3. Compute WER/CER
    # 4. Extract financial facts
    # 5. Compute FER

---

11. Using FinASR-Bench for ASR Adaptation

The training split can also be used for domain adaptation.

For example:

from datasets import load_dataset

dataset = load_dataset(
    "wuxianchao/FinASR-Bench",
    "synthetic-en",
)

train_set = dataset["train"]
validation_set = dataset["validation"]
test_set = dataset["test"]

Possible adaptation methods include:

• full fine-tuning;
• LoRA;
• adapter tuning;
• prompt/contextual biasing;
• vocabulary adaptation;
• synthetic-to-real transfer;
• multilingual financial ASR adaptation.

For fair evaluation, the test split should not be used during model adaptation or hyperparameter selection.

---

12. Real vs. Synthetic Evaluation

The real and synthetic subsets serve complementary purposes.

Real-world subsets

The real-* configurations are intended to evaluate performance under naturally occurring speech conditions.

They are useful for studying:

• realistic acoustic variation;
• spontaneous speech;
• speaker variation;
• real transcription errors;
• domain transfer.

Synthetic subsets

The synthetic-* configurations provide controlled coverage of specific financial phenomena.

They are useful for:

• targeted stress testing;
• controlled ablation studies;
• rare financial terminology;
• numerical errors;
• negation and comparison errors;
• multi-entity binding;
• long-context financial information.

We recommend reporting results on both types when possible.

---

13. Cross-Domain Experiments

The six configurations also support cross-domain adaptation experiments.

Examples include:

Synthetic-EN -> Real-EN
Real-EN      -> Synthetic-EN

Synthetic-ZH -> Real-ZH
Real-ZH      -> Synthetic-ZH

Synthetic-JA -> Real-JA
Real-JA      -> Synthetic-JA

These experiments can be used to investigate whether synthetic financial speech improves performance on real-world financial ASR, and whether adaptation to real-world data generalizes to controlled financial challenges.

---

14. Example: Inspecting Financial Annotations

from datasets import load_dataset
from pprint import pprint

dataset = load_dataset(
    "wuxianchao/FinASR-Bench",
    "synthetic-en",
)

sample = dataset["train"][0]

print("ID:")
print(sample["id"])

print("\nTranscript:")
print(sample["text"])

print("\nCategory:")
print(sample["category"])

print("\nCritical tokens:")
pprint(sample["critical_tokens"])

print("\nFinancial facts:")
pprint(sample["financial_facts"])

print("\nDerived constraints:")
pprint(sample["derived_constraints"])

print("\nRelations:")
pprint(sample["relations"])

---

15. Example: Category-Level Evaluation

FinASR-Bench supports challenge-specific evaluation.

For example:

categories = sorted(
    set(dataset["test"]["category"])
)

for category in categories:
    subset = dataset["test"].filter(
        lambda x: x["category"] == category
    )

    print(category, len(subset))

This allows reporting results such as:

Category	WER/CER	FER
Numbers / Currency / Percentage	...	...
Financial Terms	...	...
Financial Entity Relations	...	...
Negation / Condition / Comparison	...	...
Long-Context Multi-Entity	...	...

Such analysis can reveal failure modes that are hidden by aggregate WER/CER.

---

16. Dataset Statistics

Approximate average text lengths and financial-fact densities are:

Configuration	Avg. Text Length	Avg. Financial Facts
Real-EN	39.56	2.46
Real-ZH	21.12	0.73
Real-JA	55.69	1.51
Synthetic-EN	41.95	2.29
Synthetic-ZH	76.07	2.55
Synthetic-JA	108.98	2.30

The statistics illustrate that FinASR-Bench covers substantially different linguistic and financial-information densities across languages and data sources.

---

17. Intended Uses

FinASR-Bench is intended for research on:

• financial-domain ASR;
• multilingual ASR;
• domain adaptation;
• LoRA and parameter-efficient ASR adaptation;
• synthetic speech for ASR training;
• financial entity recognition;
• numerical speech recognition;
• semantic ASR evaluation;
• structured ASR evaluation;
• financially critical error analysis;
• synthetic-to-real generalization;
• robustness evaluation.

---

18. Limitations

FinASR-Bench should not be interpreted as covering all financial speech conditions.

Important limitations include:

1. The synthetic subsets depend on the characteristics of the TTS system and reference speakers used during synthesis.
2. Synthetic speech does not fully reproduce the acoustic variability of real financial conversations.
3. The distribution of financial concepts is controlled and may differ from naturally occurring financial speech.
4. Financial-fact extraction itself can introduce errors when FER is computed automatically.
5. Different languages have different linguistic and normalization characteristics, so WER/CER and FER should be interpreted jointly.
6. The benchmark focuses on financial-information preservation and does not replace general ASR evaluation.

---

19. Data Source and Licensing

The real-world portions of FinASR-Bench are derived from existing speech resources, including:

• SPGISpeech / SPGISpeech2
• Earnings-22 / Earnings-25
• WenetSpeech
• ReazonSpeech

Synthetic speech also uses reference speakers originating from speech resources including SPGISpeech, AISHELL-1, and ReazonSpeech.

Users should review the licenses and terms of the corresponding upstream datasets before redistributing, repackaging, or using source-derived audio outside the permissions granted by those resources.

The license of this repository does not supersede the licenses or terms of upstream datasets.

---

20. FER Implementation

The implementation of Structured Financial Error Rate is available separately:

https://github.com/wuxianchao/finasr-fer

The FER repository contains the structured financial evaluation pipeline, including:

Reference / Hypothesis
        |
        v
Financial Fact Extraction
        |
        v
Canonicalization
        |
        v
Fact Alignment
        |
        v
Structured Error Computation
        |
        v
Severity-Weighted Scoring
        |
        v
FER

Please refer to the FER repository for the latest metric definition and implementation details.

---

21. Citation

If you use FinASR-Bench or Structured Financial Error Rate in your research, please cite the associated paper.

@inproceedings{wu2027finasr,
  title     = {Structured Financial Error Rate for Financial ASR},
  author    = {Wu, Xianchao and others},
  booktitle = {Proceedings of xxx},
  year      = {2027}
}

Note: Please replace the BibTeX entry above with the final publication metadata once the paper is published.

---

22. Reproducibility

For reproducible experiments, we recommend reporting:

• FinASR-Bench configuration;
• train/validation/test split;
• ASR model and checkpoint;
• decoding parameters;
• text normalization procedure;
• WER/CER implementation;
• semantic-distance model;
• FER version;
• financial-fact extraction model;
• adaptation method;
• LoRA rank and hyperparameters, if applicable.

For example:

Dataset: FinASR-Bench / Real-EN
ASR: Whisper-large-v3
Adaptation: LoRA
LoRA rank: 64
Metrics:
  - WER
  - normalized WER
  - Semantic Distance
  - FER v0.7

---

23. Contact

For questions, issues, or suggestions, please open an issue in the corresponding GitHub repository or contact the authors.

• Dataset: wuxianchao/FinASR-Bench
• FER implementation: wuxianchao/finasr-fer

Contributions and reproducible benchmark results are welcome.