hey-lemon/streaming-dictation-asr

Streaming Dictation ASR

Data for the paper "Beyond Word Error Rate: A Failure Taxonomy of Streaming Dictation ASR" (Harras Mansoor, Muhammad Wisal, Hassan Bhatti, Masood Shakir, 2026). Code: https://github.com/heylemon-ai/streaming-dictation-asr

What is here

Path	Contents
data/en-dictation/	73 scripted dictations (chat, email, notes, AI prompts) read in a quiet room, 61.6 min
data/en-dictation-noise/	17 of the scripts with a fan running, 8.7 min
data/en-dictation-whisper/	16 of the scripts whispered, 11.1 min
data/nonspeech/	10 generated no-speech clips (silence, room noise, mains hum, a noise burst), 1.8 min
data/metadata.csv	one row per recording (file_name, set, condition, duration_s, script), which pairs each WAV with its script in the dataset viewer
data/<set>/transcripts.tsv	the script each clip was read from (the reference used for most results)
data/verbatim/	the word-for-word references: a draft proposed from the systems' agreement (draft.json) and the confirmed decisions (decisions.json); paper/analysis/verbatim.ts in the code repository combines them
data/third-party-clips.tsv	the 220 clips used from FLEURS, SADA, Casablanca, VoxPopuli and OpenSLR SLR74 (set, file, source, source id or transcript). The audio is not included.
data/codeswitch/mixes.tsv	the recipe for the 400 two-language recordings (which two clips, in which order)
outputs/clips/*.tsv	every tested system's transcript, timing and error for every clip in the paper, one file per run

All audio is 16 kHz mono 16-bit WAV.

How the recordings were made

One member of the authors' team read the scripts, moving between US, South Asian, British and Irish English, often within a message. The scripts were written for the study, and the people they name are invented, except that three email scripts are signed with the name of one of the authors, who agreed to this; the scripts contain no other personal data. The speaker consented to the recordings being used for the study and released publicly. The no-speech clips are generated deterministically by scripts/make-nonspeech.ts in the code repository.

Using it

To rebuild the paper's tables, follow "Rebuild the paper's tables" in the code repository's README: copy data/ into the repository's data/ folder and outputs/clips/ into cache/clips/.

outputs/clips/ columns: set, clip, engine (system id), run, reference, transcript, errors, ref_tokens, latency_ms, error (a connection or timeout failure; dropped means the client's voice gate never sent the clip), and in the 29 October runs first_final_ms, first_final_before_end and audio_ms. Results for hosted services reflect their versions at the time of testing (October and November 2026).

Licence

The in-house recordings, references and system outputs are released under CC BY 4.0. Rows in outputs/clips/ and data/third-party-clips.tsv that come from third-party datasets (FLEURS, SADA, Casablanca, VoxPopuli, OpenSLR SLR74, EARS, VOiCES) carry those datasets' reference text, which remains under the original licences (EARS: CC BY-NC 4.0; VOiCES: CC BY 4.0). The audio of those datasets is not included.

Citation

@misc{mansoor2026beyondwer,
  title  = {Beyond Word Error Rate: A Failure Taxonomy of Streaming Dictation {ASR}},
  author = {Mansoor, Harras and Wisal, Muhammad and Bhatti, Hassan and Shakir, Masood},
  year   = {2026},
  note   = {arXiv preprint}
}