Datasets:
Test Arabic Dataset
Hard test recordings: call-centre calls (two-channel), Gulf, Levantine and other dialects, Arabic-English
code-switching, and English. The reference transcripts are in metadata.jsonl, one row per audio file.
| Field | Meaning |
|---|---|
file_name |
the audio file |
transcription |
the whole recording, segments in time order |
channels |
1 or 2; in two-channel calls each channel is one speaker |
segments |
channel, start and end (seconds) and text of each speech segment |
How the transcripts were made
- Each recording was cut at pauses into segments of at most 30 s; two-channel calls one channel at a time.
- Five commercial speech-to-text systems transcribed every segment: ElevenLabs Scribe v2, Meta Muse Voice Transcribe 1.0, Microsoft MAI-Transcribe-2, OpenAI GPT-Transcribe and xAI Grok STT 1.0.
- A ROVER vote combined them word by word; a word needed at least two systems.
- Each segment was then reviewed against all five outputs, in conversation order, by Claude (an LLM, which did not listen to the audio) and written out as the most plausible transcript, using only words the systems heard.
Conventions: dialect as spoken; Arabic in Arabic script; English words in Latin script, except loanwords the client writes in Arabic script (اوكي، واتساب، ايميل، انترنت ...); numbers as digits, grouped as spoken ("22 44 55"); e-mail addresses in written form; no sound tags.
These are machine-made references, not human-verified. Most disagreement between the systems was on very short segments (backchannels, ringing, line noise).
- Downloads last month
- 20