Skip to content

Latest commit

 

History

History
200 lines (177 loc) · 10.2 KB

RESULTS.md

File metadata and controls

200 lines (177 loc) · 10.2 KB
Raw

Results

Streaming Zipformer-Transducer (Pruned Stateless Transducer + Streaming Zipformer)

pruned_transducer_stateless7_streaming

See k2-fsa#892 for more details.

You can find a pretrained model, training logs, decoding logs, and decoding results at: https://huggingface.co/TeoWenShen/icefall-asr-csj-pruned-transducer-stateless7-streaming-230208

Number of model parameters: 75688409, i.e. 75.7M.

training on disfluent transcript

The CERs are:

decoding method chunk size eval1 eval2 eval3 excluded valid average decoding mode
fast beam search 320ms 5.39 4.08 4.16 5.4 5.02 --epoch 30 --avg 17 simulated streaming
fast beam search 320ms 5.34 4.1 4.26 5.61 4.91 --epoch 30 --avg 17 chunk-wise
greedy search 320ms 5.43 4.14 4.31 5.48 4.88 --epoch 30 --avg 17 simulated streaming
greedy search 320ms 5.44 4.14 4.39 5.7 4.98 --epoch 30 --avg 17 chunk-wise
modified beam search 320ms 5.2 3.95 4.09 5.12 4.75 --epoch 30 --avg 17 simulated streaming
modified beam search 320ms 5.18 4.07 4.12 5.36 4.77 --epoch 30 --avg 17 chunk-wise
fast beam search 640ms 5.01 3.78 3.96 4.85 4.6 --epoch 30 --avg 17 simulated streaming
fast beam search 640ms 4.97 3.88 3.96 4.91 4.61 --epoch 30 --avg 17 chunk-wise
greedy search 640ms 5.02 3.84 4.14 5.02 4.59 --epoch 30 --avg 17 simulated streaming
greedy search 640ms 5.32 4.22 4.33 5.39 4.99 --epoch 30 --avg 17 chunk-wise
modified beam search 640ms 4.78 3.66 3.85 4.72 4.42 --epoch 30 --avg 17 simulated streaming
modified beam search 640ms 5.77 4.72 4.73 5.85 5.36 --epoch 30 --avg 17 chunk-wise

Note: simulated streaming indicates feeding full utterance during decoding using decode.py, while chunk-size indicates feeding certain number of frames at each time using streaming_decode.py.

The training command was:

./pruned_transducer_stateless7_streaming/train.py \
  --feedforward-dims  "1024,1024,2048,2048,1024" \
  --world-size 8 \
  --num-epochs 30 \
  --start-epoch 1 \
  --use-fp16 1 \
  --exp-dir pruned_transducer_stateless7_streaming/exp_disfluent_2_pad30 \
  --max-duration 375 \
  --transcript-mode disfluent \
  --lang data/lang_char \
  --manifest-dir /mnt/host/corpus/csj/fbank \
  --pad-feature 30 \
  --musan-dir /mnt/host/corpus/musan/musan/fbank

The simulated streaming decoding command was:

for chunk in 64 32; do
    for m in greedy_search fast_beam_search modified_beam_search; do
        python pruned_transducer_stateless7_streaming/decode.py \
            --feedforward-dims  "1024,1024,2048,2048,1024" \
            --exp-dir pruned_transducer_stateless7_streaming/exp_disfluent_2_pad30 \
            --epoch 30 \
            --avg 17 \
            --max-duration 350 \
            --decoding-method $m \
            --manifest-dir /mnt/host/corpus/csj/fbank \
            --lang data/lang_char \
            --transcript-mode disfluent \
            --res-dir pruned_transducer_stateless7_streaming/exp_disfluent_2_pad30/github/sim_"$chunk"_"$m" \
            --decode-chunk-len $chunk \
            --pad-feature 30 \
            --gpu 0
    done
done

The streaming chunk-wise decoding command was:

for chunk in 64 32; do
    for m in greedy_search fast_beam_search modified_beam_search; do
        python pruned_transducer_stateless7_streaming/streaming_decode.py \
            --feedforward-dims  "1024,1024,2048,2048,1024" \
            --exp-dir pruned_transducer_stateless7_streaming/exp_disfluent_2_pad30 \
            --epoch 30 \
            --avg 17 \
            --max-duration 350 \
            --decoding-method $m \
            --manifest-dir /mnt/host/corpus/csj/fbank \
            --lang data/lang_char \
            --transcript-mode disfluent \
            --res-dir pruned_transducer_stateless7_streaming/exp_disfluent_2_pad30/github/stream_"$chunk"_"$m" \
            --decode-chunk-len $chunk \
            --gpu 2 \
            --num-decode-streams 40
    done
done

training on fluent transcript

The CERs are:

decoding method chunk size eval1 eval2 eval3 excluded valid average decoding mode
fast beam search 320ms 4.19 3.63 3.77 4.43 4.09 --epoch 30 --avg 12 simulated streaming
fast beam search 320ms 4.06 3.55 3.66 4.70 4.04 --epoch 30 --avg 12 chunk-wise
greedy search 320ms 4.22 3.62 3.82 4.45 3.98 --epoch 30 --avg 12 simulated streaming
greedy search 320ms 4.13 3.61 3.85 4.67 4.05 --epoch 30 --avg 12 chunk-wise
modified beam search 320ms 4.02 3.43 3.62 4.43 3.81 --epoch 30 --avg 12 simulated streaming
modified beam search 320ms 3.97 3.43 3.59 4.99 3.88 --epoch 30 --avg 12 chunk-wise
fast beam search 640ms 3.80 3.31 3.55 4.16 3.90 --epoch 30 --avg 12 simulated streaming
fast beam search 640ms 3.81 3.34 3.46 4.58 3.85 --epoch 30 --avg 12 chunk-wise
greedy search 640ms 3.92 3.38 3.65 4.31 3.88 --epoch 30 --avg 12 simulated streaming
greedy search 640ms 3.98 3.38 3.64 4.54 4.01 --epoch 30 --avg 12 chunk-wise
modified beam search 640ms 3.72 3.26 3.39 4.10 3.65 --epoch 30 --avg 12 simulated streaming
modified beam search 640ms 3.78 3.32 3.45 4.81 3.81 --epoch 30 --avg 12 chunk-wise

Note: simulated streaming indicates feeding full utterance during decoding using decode.py, while chunk-size indicates feeding certain number of frames at each time using streaming_decode.py.

The training command was:

./pruned_transducer_stateless7_streaming/train.py \
  --feedforward-dims  "1024,1024,2048,2048,1024" \
  --world-size 8 \
  --num-epochs 30 \
  --start-epoch 1 \
  --use-fp16 1 \
  --exp-dir pruned_transducer_stateless7_streaming/exp_fluent_2_pad30 \
  --max-duration 375 \
  --transcript-mode fluent \
  --lang data/lang_char \
  --manifest-dir /mnt/host/corpus/csj/fbank \
  --pad-feature 30 \
  --musan-dir /mnt/host/corpus/musan/musan/fbank

The simulated streaming decoding command was:

for chunk in 64 32; do
    for m in greedy_search fast_beam_search modified_beam_search; do
        python pruned_transducer_stateless7_streaming/decode.py \
            --feedforward-dims  "1024,1024,2048,2048,1024" \
            --exp-dir pruned_transducer_stateless7_streaming/exp_fluent_2_pad30 \
            --epoch 30 \
            --avg 12 \
            --max-duration 350 \
            --decoding-method $m \
            --manifest-dir /mnt/host/corpus/csj/fbank \
            --lang data/lang_char \
            --transcript-mode fluent \
            --res-dir pruned_transducer_stateless7_streaming/exp_fluent_2_pad30/github/sim_"$chunk"_"$m" \
            --decode-chunk-len $chunk \
            --pad-feature 30 \
            --gpu 1
    done
done

The streaming chunk-wise decoding command was:

for chunk in 64 32; do
    for m in greedy_search fast_beam_search modified_beam_search; do
        python pruned_transducer_stateless7_streaming/streaming_decode.py \
            --feedforward-dims  "1024,1024,2048,2048,1024" \
            --exp-dir pruned_transducer_stateless7_streaming/exp_fluent_2_pad30 \
            --epoch 30 \
            --avg 12 \
            --max-duration 350 \
            --decoding-method $m \
            --manifest-dir /mnt/host/corpus/csj/fbank \
            --lang data/lang_char \
            --transcript-mode fluent \
            --res-dir pruned_transducer_stateless7_streaming/exp_fluent_2_pad30/github/stream_"$chunk"_"$m" \
            --decode-chunk-len $chunk \
            --gpu 3 \
            --num-decode-streams 40
    done
done

Comparing disfluent to fluent

$$ \texttt{CER}^{f}_d = \frac{\texttt{sub}_f + \texttt{ins} + \texttt{del}_f}{N_f} $$

This comparison evaluates the disfluent model on the fluent transcript (calculated by disfluent_recogs_to_fluent.py), forgiving the disfluent model's mistakes on fillers and partial words. It is meant as an illustrative metric only, so that the disfluent and fluent models can be compared.

decoding method chunk size eval1 (d vs f) eval2 (d vs f) eval3 (d vs f) excluded (d vs f) valid (d vs f) decoding mode
fast beam search 320ms 4.54 vs 4.19 3.44 vs 3.63 3.56 vs 3.77 4.22 vs 4.43 4.22 vs 4.09 simulated streaming
fast beam search 320ms 4.48 vs 4.06 3.41 vs 3.55 3.65 vs 3.66 4.26 vs 4.7 4.08 vs 4.04 chunk-wise
greedy search 320ms 4.53 vs 4.22 3.48 vs 3.62 3.69 vs 3.82 4.38 vs 4.45 4.05 vs 3.98 simulated streaming
greedy search 320ms 4.53 vs 4.13 3.46 vs 3.61 3.71 vs 3.85 4.48 vs 4.67 4.12 vs 4.05 chunk-wise
modified beam search 320ms 4.45 vs 4.02 3.38 vs 3.43 3.57 vs 3.62 4.19 vs 4.43 4.04 vs 3.81 simulated streaming
modified beam search 320ms 4.44 vs 3.97 3.47 vs 3.43 3.56 vs 3.59 4.28 vs 4.99 4.04 vs 3.88 chunk-wise
fast beam search 640ms 4.14 vs 3.8 3.12 vs 3.31 3.38 vs 3.55 3.72 vs 4.16 3.81 vs 3.9 simulated streaming
fast beam search 640ms 4.05 vs 3.81 3.23 vs 3.34 3.36 vs 3.46 3.65 vs 4.58 3.78 vs 3.85 chunk-wise
greedy search 640ms 4.1 vs 3.92 3.17 vs 3.38 3.5 vs 3.65 3.87 vs 4.31 3.77 vs 3.88 simulated streaming
greedy search 640ms 4.41 vs 3.98 3.56 vs 3.38 3.69 vs 3.64 4.26 vs 4.54 4.16 vs 4.01 chunk-wise
modified beam search 640ms 4 vs 3.72 3.08 vs 3.26 3.33 vs 3.39 3.75 vs 4.1 3.71 vs 3.65 simulated streaming
modified beam search 640ms 5.05 vs 3.78 4.22 vs 3.32 4.26 vs 3.45 5.02 vs 4.81 4.73 vs 3.81 chunk-wise
average (d - f) 0.43 -0.02 -0.02 -0.34 0.13