What happened
The Miracle Channel Association runs CJIL-DT, a television station in Lethbridge, Alberta. It captions its live news with EEG's automatic speech-to-text software, with no human captioner. From 1 March 2020 the CRTC required live English-language captions to score at least 98 under the NER model. In 2019 a group of independent stations that included the Miracle Channel had told the CRTC that voice recognition captioning could struggle to meet it. Pablo Romero-Fresco, who developed the model, and Nazaret Fresno analyzed the broadcasters' reports to the CRTC. They report that EEG's cloud version reached 98% in 8/31 samples in 2019-2020. The station then moved to Lexi Local, on a local server, and a staff member compiled glossaries with phonetic spellings. In 2020-2021 only 1/24 samples fell below 98%. The station chose its samples, and an evaluator that it engaged scored them. Its report for 2023-2024 lists scores from 98.49 to 99.38. Across the broadcasters studied, 60% of automatic samples missed the threshold. The authors note that unaided automatic captions still fell short of it.
Sources
-
Peer-reviewed study, by authors with no tie to the system.
Romero-Fresco P, Fresno N. The accuracy of automatic and human live captions in English. Linguistica Antverpiensia, New Series: Themes in Translation Studies 22 (2023): 114-133, section 3.4 (case study 4: live TV captions in Canada). Published 13 December 2023. doi:10.52034/lans-tts.v22i.774.
for the first time, broadcasters such as Jim Pattison Broadcast Group and The Miracle Channel Associations consistently reach accuracy rates above the required threshold in 2020-2021. Impressively, of the 55 samples analysed by these two broadcasters in 2020-2021, 52 reached 98%.
In 2019-2020, their cloud-based version of the automatic captioning software EEG reached the 98% accuracy rate in 8/31 samples and 10/31 scored below 97.5%. For the 2020-2021 round, they decided to use the local server of the software (called Lexi Local) and to have a dedicated member of the staff compile glossaries with their phonetic spellings to improve the accuracy of the captions. This enabled them to reach the NER threshold consistently, with only 1/24 samples below par but pretty close to the minimum accuracy rate (97.92%).
They are also more accurate than automatic captions, which feature more samples below 98% accuracy and fewer samples exceeding a 99% accuracy rate (60% of the automatic samples did not reach the minimum threshold and only 1% featured an excellent accuracy).
-
Evaluation report, by an evaluator that the makers or users commissioned.
The Miracle Channel Association (CJIL-DT). NER98 Accuracy Standards Summary for Miracle Channel (CJIL-DT) 2020-21 and Live-Captioning NER Score Report for 2020-2021, with scores by a certified NER evaluator that the broadcaster engaged. Annual report to the CRTC under Broadcasting Regulatory Policy CRTC 2019-308. Archived copy of https://crtc.gc.ca/eng/BCASTING/ann_rep/miracle_cc2021.pdf.
We are maintaining consistent NER98 accuracy through the use of an on-premise speech-to-text service called Lexi Local, from EEG Enterprises.
-
Regulatory document, by authors with no tie to the system. Context or contrary evidence.
Canadian Radio-television and Telecommunications Commission. Broadcasting Regulatory Policy CRTC 2019-308: English-language closed captioning mandatory quality standards relating to the accuracy rate for live television programming. Ottawa, 30 August 2019. Archived copy of https://crtc.gc.ca/eng/archive/2019/2019-308.htm. Appendix 1, Quality Standard 3.
Beginning 1 March 2020, broadcasters must reach an accuracy rate of at least 98 for English-language closed captioning of live television programming
Further, E-ILTS acknowledged that voice recognition software, which is sometimes used by independent television stations and allows for costs savings, is improving year by year, but noted that based on preliminary testing, voice recognition-based captioning could have difficulty meeting the proposed NER 98 accuracy standard in the short term.
Jim Pattison Broadcast Group Limited Partnership, operating in Kamloops and Prince George, British Columbia, and Medicine Hat, Alberta; Stingray Digital Group Inc. (now Stingray Group Inc.) in Lloydminster, Alberta; Thunder Bay Electronics Limited in Thunder Bay, Ontario; The Miracle Channel Association in Lethbridge, Alberta;
-
Report, by the organization that used it. Context or contrary evidence.
The Miracle Channel Association (CJIL-DT). Summary of The Miracle Channel Association (CJIL-DT) Efforts to Comply with BRP 2019-308 and NER98 Accuracy Standards, with the Live-Captioning NER Score Report for 2019-2020. Annual report to the CRTC, 30 November 2020. Archived copy of https://crtc.gc.ca/eng/BCASTING/ann_rep/miracle_cc2020.pdf.
The review process involved reviewing the video segment that was captioned and fixing all the mistakes that were presented in the exported log file and then uploading that back into the cloud service to help improve the AI and its dictionary in the hope of improving the accuracy moving forward.
Details
AI system
- Workflow
-
EEG's speech-to-text service captioned the station's live news with no human captioner. In 2020-2021 the station moved from the cloud version of the software to Lexi Local, on a local server. A staff member compiled glossaries of words with their phonetic spellings. Staff adjusted the glossary from the errors that the certified evaluator found.
Follow-up
- Conditions for replication
-
Live English-language captions must score at least 98 under the Canadian NER guidelines, as calculated by certified evaluators. Two programs with live content must be scored each month. Automatic captions need effective human help, such as glossaries, to reach the threshold regularly.
Entry details
- Sector
-
Broadcasting / live closed captions for deaf and hard-of-hearing viewers.
- Purpose
-
EEG Enterprises built Lexi Local, an on-premises version of its Lexi captioning service, to produce live captions automatically from speech.
31 more questions not yet checked
model, release, provider, role, configuration, inputs and outputs, autonomy, human oversight, AI contribution, beneficiaries, reach, distribution and access, duration, every measurement question, every money question, ownership and control, dependencies, cross-border connections, geopolitical implications, replication attempts, availability, institutional decision, grounds, action and result, success criterion, every evidence question, observed or intended, alternative assessments, related entries, other identifiers, category mappings.
Cite this record
Peter H. Vartanian. (2026). Lexi Local captions meet Canada's live accuracy standard in Lethbridge (AIBD-1147). Artificial Intelligence Benefits Database, edition of 1 November 2026. https://petervartanian.xyz/aibd/records/AIBD-1147.html (archived at https://doi.org/10.7910/DVN/AADSII, version 7.0)