Latest in AI
OhanASR for Armenian: Choose Recorded Transcription or Live Captions
Metric-AI released two Armenian speech models. Compare their accuracy, access terms and latency before choosing one for a recorded or live workflow.
Metric-AI's November 20 OhanASR release gives Armenian speech projects two checkpoints rather than one: a model for recorded speech and a streaming variant for text that appears while someone is speaking.
If you are transcribing an archive, start with the offline model. If you need captions during an event, evaluate the streaming model separately. The thing to compare is not just whether both produce text. It is how quickly the text arrives, what it gets wrong, and whether the workflow can correct those errors.
Check access before planning deployment
Both checkpoints use the Creative Commons Attribution-NonCommercial 4.0 license. Hugging Face lists the model cards publicly, but accessing the files requires signing in and agreeing to share account information. Neither model lists a hosted Inference Provider on its card.
Treat them as candidates for research or other noncommercial evaluation, not drop-in commercial production endpoints. Teams building a paid transcription product should clarify the model rights before investing in implementation. Download access does not answer the licensing question.
Start with a short recorded clip
The offline model card documents Linux with Python 3.10 or 3.11 and pinned dependencies, including the audio library needed by fairseq2. The checkpoint is fairseq2 format and uses the omnilingual-asr loader with the bundled helper, not a Transformers AutoModel call.
Take a short, public or approved WAV file and run the card's offline example. Save the checkpoint revision, dependency versions and device with your result. The helper resamples to mono at 16 kHz and can run float32 on CPU. Measure processing time on the device you actually intend to use.
For a longer recording, the helper looks for quieter places to split audio and keeps each segment within the model's 40-second limit. Check the text either side of a cut. A word sliced in two can produce an error even if the rest of the recording is transcribed well.
Don't confuse lookahead with caption latency
The streaming model card describes one-second audio chunks and 0.3 seconds of lookahead beyond each chunk. That 0.3 seconds is not the time between a person's first word and the corresponding caption appearing on screen. Your application also has to receive enough audio, run inference and render the result.
Start with the card's file-based callback example. It emits text for successive chunks, but the documented transcribe helper encodes the supplied audio in one pass and uses more memory as input gets longer. A successful replay is not evidence of a bounded live microphone service. Test the live capture path and long-session memory separately before promising either.
Time from the point a person speaks to the point the text becomes visible. For an event caption view, also test pauses and fast speech. Measurements from short, clean files miss the delays an audience notices.
Compare the same scoring rules
Metric-AI reports combined word error rates of 16.12% for offline and 19.07% for streaming in its benchmark. The cards clarify that this column already normalizes numbers. Adding text normalization lowers the scores to 6.96% and 8.57%, respectively. Those columns are two ways of scoring output, not two model upgrades.
Keep a strict transcript score alongside a normalized one. For subtitles, punctuation and spelling affect the edit workload; for search, a variant may still be usable. If you strip differences before scoring, make sure you have not also stripped the errors people care about.
The publisher flags code-switching and unclear audio as limitations. Questions containing English product names inside Armenian sentences deserve their own slice. Count how often a name changes spelling or disappears, even if the sentence remains understandable.
Pick a workflow, then pick a winner
For archived talks, compare editing time as well as error rates. For live captions, compare latency and the frequency of missing names, numbers and negation. Review noisy recordings and spontaneous conversation separately from scripted speech.
An offline pass after an event can be a useful second test, but do not assume it always repairs the earlier output. Keep draft captions and the reviewed transcript separate so readers know which version they are looking at.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI teams explain model releases and real integration tradeoffs through technical writing developers can use.