Skip to main content

DeafBench

Evaluate what ASR metrics miss.

DeafBench is an open-source benchmark for measuring AI caption failures that matter to Deaf and hard-of-hearing users.

Traditional ASR benchmark:

Model A: 7.8% WER
Model B: 8.1% WER

DeafBench:

Model A: 72% critical information preserved
Model B: 96% critical information preserved

Two models can have similar Word Error Rate (WER) while producing very different accessibility outcomes. DeafBench highlights critical information loss (names, numbers, negations, dates, technical terms) and non-speech annotations that traditional WER obscures.


Quickstart

Installation

pip install deafbench

Or install locally for development:

git clone https://github.com/488315/DeafBench.git
cd DeafBench
pip install -e .

Usage

1. Compare predictions against reference captions:

deafbench compare examples/references.jsonl examples/model-a.jsonl

Output:

DeafBench v0.1

Samples: 3

WER                         8.4%
Critical Information       91.2%
Non-Speech Information     62.5%
Speaker Attribution        87.0%
Median Latency             1.4s

⚠ 2 critical-information failures detected

2. Generate a Markdown evaluation report:

deafbench report examples/references.jsonl examples/model-a.jsonl --output report.md

Input JSONL Schema

references.jsonl

{
  "id": "sample-001",
  "text": "John Doe needs 25 milligrams on Friday.",
  "critical": ["John Doe", "25 milligrams", "Friday"],
  "sounds": ["[alarm]"],
  "speaker": "Speaker 1"
}

predictions.jsonl

{
  "id": "sample-001",
  "text": "Guy needs 20 milligrams on Friday.",
  "latency_ms": 820,
  "speaker": "Speaker 1"
}

License

Apache License 2.0

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

deafbench-0.1.1.tar.gz (14.1 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

deafbench-0.1.1-py3-none-any.whl (17.3 kB view details)

Uploaded Python 3

File details

Details for the file deafbench-0.1.1.tar.gz.

File metadata

  • Download URL: deafbench-0.1.1.tar.gz
  • Upload date:
  • Size: 14.1 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for deafbench-0.1.1.tar.gz
Algorithm Hash digest
SHA256 3207b00f92d4fb8b88c29894ea8fdf340fc3498d854b7c903e99d62816647eac
MD5 28759e795f80f1b11a63ddf9f8f52db4
BLAKE2b-256 f17cafe185c7ae00c6174bdd777a94ac4940443bd42a73a1ab6019089ab8433a

See more details on using hashes here.

Provenance

The following attestation bundles were made for deafbench-0.1.1.tar.gz:

Publisher: release.yml on 488315/DeafBench

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file deafbench-0.1.1-py3-none-any.whl.

File metadata

  • Download URL: deafbench-0.1.1-py3-none-any.whl
  • Upload date:
  • Size: 17.3 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for deafbench-0.1.1-py3-none-any.whl
Algorithm Hash digest
SHA256 201726347cdaf4b0784cbfc282cbda26c10c651af97e3c7ac5ebf076f06e4469
MD5 5b49cd770ce167163fdd8de662a52093
BLAKE2b-256 9b3821d08f0fe4549d499619a7003efea0c5dad18dff67eecd7ad807df8c8ff2

See more details on using hashes here.

Provenance

The following attestation bundles were made for deafbench-0.1.1-py3-none-any.whl:

Publisher: release.yml on 488315/DeafBench

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page