license: cc-by-nc-4.0
extra_gated_heading: Acknowledge license to access the dataset
extra_gated_description: >-
To request access, please review and accept the terms below. Our team may take
2–3 business days to process your request.
extra_gated_button_content: Request access
extra_gated_prompt: >-
This data is licensed under CC-BY-NC. Those terms notwithstanding, by
downloading this data you agree that: You will use this data only for the
purposes of testing transcription quality; You will not share or redistribute
this data; and At the earliest of your completion of use of the data for the
purpose described above, or 10 days, you will securely delete the downloaded
data and any copies you may have created or allowed to be created.
extra_gated_fields:
Country: country
Company/Affiliation: text
Role: text
I agree to use this dataset for non-commercial purposes only: checkbox
Dataset Card for μ-Bench (Leaderboard | Code)
μ-Bench is a multilingual transcription benchmark built from real customer-service phone conversations.
Dataset Details
Dataset Description
Most public ASR benchmarks are either English-only or built from read speech in quiet studios. μ-Bench fills that gap: real phone-call audio, five languages, and metrics that go beyond Word Error Rate to distinguish meaning-changing errors from surface-level ones.
The calls are scripted interactions with an AI banking agent built on Sierra's voice platform — checking card status, confirming case codes, providing personal details, disputing transactions, and requesting credit-limit increases.
Callers used their own phones from their own environments, producing realistic background noise, disfluencies, emotional variation, interruptions, and diverse speaking styles.
The call audio is processed and clipped into individual user utterances.
Dataset Structure
The dataset contains 250 conversations totaling 4,270 utterances (~5.1 hours of audio):
| Locale | Language | Utterances |
|---|---|---|
| en-US | English (US) | 817 |
| es-MX | Spanish (Mexico) | 792 |
| tr-TR | Turkish | 846 |
| vi-VN | Vietnamese | 975 |
| zh-CN | Chinese (Mandarin) | 840 |
Audio files follow:
<locale>/conv-<N>-turn-<M>.wav
Each utterance has a row in metadata.jsonl:
| Field | Type | Description |
|---|---|---|
| file_name | string | Relative audio path |
| locale | string | BCP-47 locale tag |
| conversation_id | int | Conversation identifier |
| turn_index | int | Turn index (0-based) |
| duration_sec | float | Duration in seconds |
| transcript | string | Clean verbatim transcript |
There is a single split. Evaluation uses the full dataset with conversation-level bootstrap resampling.
Evaluation Artifacts
In addition to the dataset, the prompts used for LLM-based normalization and error classification (see our Github for more details) are also available for download as part of this repository.
More Information
- Leaderboard: https://research.sierra.ai/mubench
- Code & submissions: https://github.com/sierra-research/mu-bench