doing something
https://huggingface.co/spaces/Yehor/see-asr-outputs Made a simple space to see ASR outputs generated by evaluation scripts #python #speech
Remind, there's https://huggingface.co/spaces/Yehor/evaluate-asr-outputs for calculating ASR metrics from these JSONL files
huggingface.co
Evaluate ASR outputs - a Hugging Face Space by Yehor
Calculate WER/CER values from JSONL files made by ASR models
doing something
#daily_pain
This pain was converted to a working https://huggingface.co/spaces/Yehor/evaluate-asr-outputs/blob/main/Dockerfile file that runs the latest Python (3.13.2) on Debian image
doing something
https://github.com/egorsmkv/eerie-yolo11 #rust #cv
With help of the author, now it works!
Forwarded from Yehor Smoliakov
doing something
Data Hoarding in Action: - 7.4k hours of audio; - 10.7 million of utterances; - 2 GPU - 3090; - batch size d0 = 60, d1 = 3400; - sphn (rust-based library to load audio to numpy ndarray); - RTF: ~0.0044 (data should be done in ~16 hours). #python #rust
Continuing with https://github.com/efeslab/LiteASR
Interesting that I can set maximum batch_size=320, when wav2vec2 allowed me batch_size=3400
Interesting that I can set maximum batch_size=320, when wav2vec2 allowed me batch_size=3400
doing something
Data Hoarding in Action: - 7.4k hours of audio; - 10.7 million of utterances; - 2 GPU - 3090; - batch size d0 = 60, d1 = 3400; - sphn (rust-based library to load audio to numpy ndarray); - RTF: ~0.0044 (data should be done in ~16 hours). #python #rust
Ok, my calculator was wrong
It's 10M of rows
It's 10M of rows
👍1
doing something
Added ability to calculate WER/CER metrics per each row for better visual debugging #speech #ai
Results are showed in a data frame, Gradio has many features with it:
https://x.com/gradio/status/1904784469300830598?s=46&t=7jwH29MvU0R301CgvqVBYw
#ml
https://x.com/gradio/status/1904784469300830598?s=46&t=7jwH29MvU0R301CgvqVBYw
#ml
Forwarded from Feed-Master
doing something
Added ability to calculate WER/CER metrics per each row for better visual debugging #speech #ai
Added culculation of Levenshtein distance, it's easier to compute
Used: https://github.com/ion-elgreco/polars-distance
#rust
Used: https://github.com/ion-elgreco/polars-distance
#rust
doing something
Added culculation of Levenshtein distance, it's easier to compute Used: https://github.com/ion-elgreco/polars-distance #rust
Btw, it uses https://github.com/rapidfuzz/rapidfuzz-rs/ under the hood
GitHub
GitHub - rapidfuzz/rapidfuzz-rs: Rapid fuzzy string matching in Rust using various string metrics
Rapid fuzzy string matching in Rust using various string metrics - GitHub - rapidfuzz/rapidfuzz-rs: Rapid fuzzy string matching in Rust using various string metrics