icefall_pruned_transducer_stateless7_streaming_et

Icefall streaming ASR model for Estonian

This is a streaming end-to-end transducer model for Estonian, trained using Icefall

It is trained on around 800 h of manually transcribed speech from various domains and on about 2500 h of automatically transcribed speech from Estonian TV (mainly news and talkshows)

Serving

To use it on a server for browser-based ASR:

• Install Sherpa

• Clone this model locally:

  git lfs install
  git clone https://huggingface.co/TalTechNLP/icefall_pruned_transducer_stateless7_streaming_et
  

• Set SHERPA_ROOT_DIR to the sherpa root directory

• Start serving on port 6006:

  sherpa-online-websocket-server --use-gpu=false --decode-chunk-size=32  \
    --encoder-model=icefall_pruned_transducer_stateless7_streaming_et/exp/1d/encoder_jit_trace.pt \
    --decoder-model=icefall_pruned_transducer_stateless7_streaming_et/exp/1d/decoder_jit_trace.pt \
    --joiner-model=icefall_pruned_transducer_stateless7_streaming_et/exp/1d/joiner_jit_trace.pt  \
    --tokens=icefall_pruned_transducer_stateless7_streaming_et/data/lang_bpe_1000/tokens.txt \
    --doc-root=${SHERPA_ROOT_DIR}/sherpa/bin/web --decoding-method=modified_beam_search
  

• Open in browser: http://localhost:6006 (also works via ssh tunnel) and go to "Streaming-Record" tab

• Click "Connect" and then "Streaming-Record" button, and start talking