[Need help] How to realize Syllable-level Voice Recognition with sherpa-onnx Open Vocabulary Keyword Spotting

[diyism]

I've been always trying to use Sherpa to implement syllable-level speech recognition (1.use a few pinyins to detect hotword directly; 2.or send a long sequence of pinyins to a LLM(gpt or claude) to convert it into the most appropriate Chinese sentence) (https://github.com/k2-fsa/sherpa-ncnn/issues/177)

I found that you released the sherpa-onnx Open Vocabulary Keyword Spotting at 2024-02(https://k2-fsa.github.io/sherpa/onnx/kws/pretrained_models/index.html)

So I imagine I can utilize it to realize syllable-level speech recognition: I've modified the sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/keywords.txt into:

j iǎng @jiang3
y ǒu @you3
b ó @bo2

h uí @hui2
d á  @da2
q ǐng @qing3
g ài @gai4
g ē  @ge1

Tested it with the AHPUymhd's code(https://github.com/k2-fsa/sherpa-onnx/issues/760), specify the sound_files = ["./sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/4.wav"], (jiang3 you3 bo2 bei4 pai1 dao4 ...) the output: jiang3/bo2/bo2

I understand that the "bo2 bo2"(伯伯 uncle) is a more frequently used word than "you3 bo2"(but the Syllable-level Voice Recognition needs to be future-proof and can recognize any new word in the future into pinyins). So I very carefully split the ./sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/4.wav to ensure each WAV file contains only one syllable:

$ sox 4.wav jiang3.wav trim 0.4 0.33
$ sox 4.wav you3.wav trim 0.77 0.2
$ sox 4.wav bo2.wav trim 1.05 0.25

Now, if I run the python code with "sound_files = ["./sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/jiang3.wav"]", it can correctly output "jiang3", and if I run it with "you3.wav", it can output "you3", if I run it with "bo2.wav", it can say "bo2", It's perfect, even if there're other interfering pinyins(h uí, d á, ...) in keywords.txt file(I am dreaming of adding all 1300 pinyins into it).

So, I guess the sherpa-onnx Open Vocabulary Keyword Spotting is fully capable of perfectly recognizing Chinese mono-syllables, but a method is needed to segment each syllable. Maybe something like silero-vad can do it.

Any idea ?

@danpovey @csukuangfj @pkufool @marcoyang1998

[pkufool]

What system do you want, can you clarify it in details. @diyism

After reading the issue you posted before, I know what you want. Actually, I already have a model modeling with pinyin in my machine, will push to huggingface.

[diyism]

Thank you, I am eager to test it.

[diyism]

What a pity, I can't find a realtime VAD for mandarin Syllables. https://github.com/linto-ai/whisper-timestamped https://github.com/readbeyond/aeneas (needs wav + text) https://modelscope.cn/models/iic/speech_timestamp_prediction-v1-16k-offline/summary (needs wav + text) https://github.com/snakers4/silero-vad

If there's a VAD that can detect the beginning of every mandarin syllables(https://courses.washington.edu/chin342/ipa/syllables.html), I can truncate every beginning 0.2 seconds of each syllable as a wav file and send it to Sherpa-ONNX Open Vocabulary Keyword Spotting for processing, as Sherpa-ONNX kws is already able to accurately recognize individual syllable wav files.

[pkufool]

@diyism Please follow the progress of this PR https://github.com/k2-fs/icefall/pull/1662 . It adds models trained with pinyin.

The recognized results look like this, I think they are what you want.

DEV_T0000000000_S00000: ref=['duì', 'wǒ', 'zuò', 'le', 'jiè', 'shào', 'a', 'nà', 'me', 'wǒ', 'xiǎng', 'shuō', 'de', 'shì', 'ne', 'dà', 'jiā', 'rú', 'guǒ', 'duì', 'wǒ', 'de', 'yán', 'jiū', 'gǎn', 'xìng', 'qù', 'ne', 'ń']
DEV_T0000000000_S00000: hyp=['duì', 'wǒ', 'zuò', 'le', 'jiè', 'shào', 'nà', 'me', 'wǒ', 'xiǎng', 'shuō', 'de', 'shì', 'dà', 'jiā', 'rú', 'guǒ', 'duì', 'wǒ', 'de', 'yán', 'jiū', 'gǎn', 'xìng', 'qù']
DEV_T0000000001_S00000: ref=['zhòng', 'diǎn', 'ne', 'xiǎng', 'kàn', 'sān', 'gè', 'wèn', 'tí', 'shǒu', 'xiān', 'ne', 'jiù', 'shì', 'zhè', 'yī', 'lún', 'quán', 'qiú', 'jīn', 'róng', 'dòng', 'dàng', 'de', 'biǎo', 'xiàn']
DEV_T0000000001_S00000: hyp=['zhòng', 'diǎn', 'xiǎng', 'tán', 'sān', 'gè', 'wèn', 'tí', 'shǒu', 'xiān', 'jiù', 'shì', 'zhè', 'yī', 'lún', 'quán', 'qiú', 'jīn', 'róng', 'dòng', 'dàng', 'de', 'biǎo', 'xiàn']
DEV_T0000000002_S00000: ref=['shēn', 'rù', 'dì', 'fēn', 'xī', 'zhè', 'yī', 'cì', 'quán', 'qiú', 'jīn', 'róng', 'dòng', 'dàng', 'bèi', 'hòu', 'de', 'gēn', 'yuán']
DEV_T0000000002_S00000: hyp=['shēn', 'rù', 'dì', 'fēn', 'xī', 'zhè', 'yī', 'cì', 'quán', 'qiú', 'jīn', 'róng', 'dòng', 'dàng', 'bèi', 'hòu', 'de', 'gēn', 'yuán']
DEV_T0000000003_S00000: ref=['a', 'jiǎng', 'dì', 'yí', 'gè', 'wèn', 'tí', 'hā', 'zěn', 'me', 'lái', 'kàn', 'dài']
DEV_T0000000003_S00000: hyp=['jiǎng', 'dì', 'yí', 'gè', 'wèn', 'tí', 'zěn', 'me', 'lái', 'kàn', 'dài']

[pkufool]

What a pity, I can't find a realtime VAD for mandarin Syllables. https://github.com/linto-ai/whisper-timestamped https://github.com/readbeyond/aeneas (needs wav + text) https://modelscope.cn/models/iic/speech_timestamp_prediction-v1-16k-offline/summary (needs wav + text) https://github.com/snakers4/silero-vad

If there's a VAD that can detect the beginning of every mandarin syllables(https://courses.washington.edu/chin342/ipa/syllables.html), I can truncate every beginning 0.2 seconds of each syllable as a wav file and send it to Sherpa-ONNX Open Vocabulary Keyword Spotting for processing, as Sherpa-ONNX kws is already able to accurately recognize individual syllable wav files.

Using vad + kws is not a normal way to implement this feature. I am very surprised you have not implemented this feature in a year, did you try following our suggestions to train a model based on pinyin? Did you have some troubles training the model? Anyway, I am adding the recipe now, see the comments above.

[diyism]

I guess "VAD splitting" can prevent interference between syllables, such as "jiang3 you3 bo2" being recognized as "jiang3 bo2 bo2" (蒋伯伯), or "gai4 ge2" being recognized as "gai4 kuo4" (概括).

Additionally, VAD can avoid issues caused by continuously sliding time windows to extract segments for recognition, like in whisper-timestamped realtime recognition.

It can also prevent the problem of missing syllables: ['zhòng', 'diǎn', 'ne', 'xiǎng', 'kàn', 'sān', 'gè', 'wèn', 'tí'] missing the middle 'ne'. Human ears and VAD can clearly know that there is a syllable 'ne' in the middle.

It seems that whisperX has the best VAD, but it's not for mandarin(https://github.com/m-bain/whisperX):

Anyway, I'm looking forward to the progress of decode_pinyin.py: https://github.com/k2-fsa/icefall/pull/1662

[diyism]

whisperX's VAD is also not a syllable-level VAD, the "That's"(in the whisperX README) have 2 syllables, and whisperX is also not a real-time VAD, it's similar to the aeneas(https://github.com/readbeyond/aeneas), force alignment of pre-prepared "wav + text".

Even if whisperX supports mandarin, the "./sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/4.wav" won't be separated syllable by syllable:

The first 3 syllables of "jiang3 you3 bo2" in the 4.wav will be seperated as "jiang3 you3" and "bo2". So the whisperX's VAD can't prepare syllable-level wav files for sherpa-onnx-kws recognition just like my manually "sox trim" 4.wav file.

[diyism]

I'm wrong, the whisperX can seperate the first 2 syllables of "./sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/4.wav":

$ sox ../sherpa-onnx/sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/4.wav a.wav trim 0 0.97
$ whisperx --compute_type int8 --language zh a.wav
$ cat a.json
{"segments": [{"start": 0.537, "end": 0.822, "text": "讲有", "words": [{"word": "讲", "start": 0.537, "end": 0.801, "score": 0.923}, {"word": "有", "start": 0.801, "end": 0.822, "score": 0.591}]}], "word_segments": [{"word": "讲", "start": 0.537, "end": 0.801, "score": 0.923}, {"word": "有", "start": 0.801, "end": 0.822, "score": 0.591}], "language": "zh"}

Maybe I can use whisperX as a "realtime syllable-level speech recognizer", but it is not enough fast:

$ time whisperx --compute_type int8 --language zh a.wav
real    0m16.209s
user    0m21.419s
sys 0m5.020s

While sherpa-onnx-kws.py is very fast, but it can't recognize the second syllable:

$ sox ./sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/4.wav a.wav trim 0 1.0
$ aplay a.wav
$ python sherpa-onnx-kws.py --sound_files ./a.wav
Started!
jiang3 is detected.
Done!
./a.wav
jiang3/
----------
num_threads: 1
Wave duration: 1.000 s
Elapsed time: 0.053 s
Real time factor (RTF): 0.053/1.000 = 0.053

And the whisperx also has the interference problem between syllables:

$ sox ../sherpa-onnx/sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/4.wav a.wav trim 0 1.2
$ aplay a.wav
$ whisperx --compute_type int8 --language zh a.wav
$ cat a.json
{"segments": [{"start": 0.536, "end": 1.062, "text": "江有博", "words": [{"word": "江", "start": 0.536, "end": 0.799, "score": 0.923}, {"word": "有", "start": 0.799, "end": 1.041, "score": 0.947}, {"word": "博", "start": 1.041, "end": 1.062, "score": 0.257}]}], "word_segments": [{"word": "江", "start": 0.536, "end": 0.799, "score": 0.923}, {"word": "有", "start": 0.799, "end": 1.041, "score": 0.947}, {"word": "博", "start": 1.041, "end": 1.062, "score": 0.257}], "language": "zh"}

The first syllable's tone is wrong.

sherpa-onnx-kws with 3 wav files of single syllable:

$ sox 4.wav jiang3.wav trim 0.4 0.33
$ sox 4.wav you3.wav trim 0.77 0.2
$ sox 4.wav bo2.wav trim 1.05 0.25
$ sox 4.wav jiang3you3bo2.wav trim 0 1.3

$ python sherpa-onnx-kws.py --sound_files ./jiang3.wav
Started!
jiang3 is detected.
$ python sherpa-onnx-kws.py --sound_files ./you3.wav
Started!
you3 is detected.
$ python sherpa-onnx-kws.py --sound_files ./bo2.wav
Started!
bo2 is detected.
$ python sherpa-onnx-kws.py --sound_files ./jiang3you3bo2.wav
Started!
jiang3 is detected.
bo2 is detected.
bo2 is detected.

Currently for first 3 syllables of ./sherpa-onnx/sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/4.wav:

whisperx with 1 wav file of 3 syllables:               jiang1 you3 bo2    (wrong)
sherpa-onnx-kws with 1 wav file of 3 syllables:        jiang3 bo2  bo2    (wrong)
sherpa-onnx-kws with 3 wav files of single syllable:   jiang3 you3 bo2    (correct)

It's perfect for sherpa-onnx-kws with 3 wav files of single syllable, but I haven't found a syllable segmentation tool.

[diyism]

There was a project about "Singing voice phoneme segmentation" (syllable segmentation) 6 years ago: https://github.com/ronggong/interspeech2018_submission01

[diyism]

There's a simple project that can segment syllables 5 years ago(https://github.com/diyism/thetaOscillator-syllable-segmentation), but what a pity, it also can't split the first 2 syllables("jiang3 you3") of ./sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/4.wav :

[diyism]

I found a simple command in ~/miniconda3/bin/sherpa-onnx-keyword-spotter which was installed by building sherpa-onnx, and I've test it with:

$ cd sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01
$ sherpa-onnx-keyword-spotter --tokens=tokens.txt     --encoder=encoder-epoch-12-avg-2-chunk-16-left-64.onnx     --decoder=decoder-epoch-12-avg-2-chunk-16-left-64.onnx     --joiner=joiner-epoch-12-avg-2-chunk-16-left-64.onnx     --provider=cpu     --num-threads=8     --keywords-file=../keywords.txt --keywords-threshold=0.2 --modeling-unit=cjkchar test_wavs/4.wav

the output:

{"start_time":0.00, "keyword": "jiang3", "timestamps": [0.64, 0.68], "tokens":["j", "iǎng"]}

test_wavs/4.wav
{"start_time":0.00, "keyword": "bo2", "timestamps": [1.12, 1.16], "tokens":["b", "ó"]}

test_wavs/4.wav
{"start_time":0.00, "keyword": "bo2", "timestamps": [1.28, 1.32], "tokens":["b", "ó"]}

It seems the timestamps are not correct, according to the previous sox commands that splited 3 wav files perfectly recognized by sherpa-onnx-keyword-spotter:

$ sox 4.wav jiang3.wav trim 0.4 0.33
$ sox 4.wav you3.wav trim 0.77 0.2
$ sox 4.wav bo2.wav trim 1.05 0.25

The correct timestamps should be:

[0.40, 0.73]
[0.77, 0.97]
[1.05, 1.30]

It seems the allosaurus project (https://github.com/xinjli/allosaurus) can produce correct timestamps but it missed a syllable of "bei"(in 5 syllables of "jiang3 you3 bo2 bei4 pai1"):

$ python -m allosaurus.run -e 1.2 --lang=cmn -i 4_8000hz.wav --timestamp=True
0.510 0.045 ɕ
0.540 0.045 i
0.600 0.045 a
0.630 0.045 ŋ

0.750 0.045 j
0.780 0.045 i
0.810 0.045 o
0.870 0.045 ə

1.020 0.045 p
1.080 0.045 o

1.590 0.045 p
1.650 0.045 a
1.710 0.045 i

Is there any parameter that can improve the timestamps results? I've already tried "--modeling-unit=cjkchar", "--wenet-ctc-chunk-size=8" and "--wenet-ctc-num-left-chunks=2", but they don't seem to have much effect.

[diyism]

If I truncate the first syllables (jiang3 and you3), and lower the "--keywords-threshold" into 0.08, it can recognize the second syllable:

$ sox test_wavs/4.wav jiang3you3.wav trim 0.0 0.97
$ sherpa-onnx-keyword-spotter --tokens=tokens.txt     --encoder=encoder-epoch-12-avg-2-chunk-16-left-64.onnx     --decoder=decoder-epoch-12-avg-2-chunk-16-left-64.onnx     --joiner=joiner-epoch-12-avg-2-chunk-16-left-64.onnx     --provider=cpu     --num-threads=8     --keywords-file=../keywords.txt --keywords-threshold=0.08 jiang3you3.wav
jiang3you3.wav
{"start_time":0.00, "keyword": "jiang3", "timestamps": [0.64, 0.68], "tokens":["j", "iǎng"]}

jiang3you3.wav
{"start_time":0.00, "keyword": "you3", "timestamps": [0.84, 0.96], "tokens":["y", "ǒu"]}

But the timestamp results are still not correct.

And use "--keywords-threshold=0.08" to parse the whole test_wavs/4.wav file, the result is still the wrong "蒋伯伯":

test_wavs/4.wav
{"start_time":0.00, "keyword": "jiang3", "timestamps": [0.64, 0.68], "tokens":["j", "iǎng"]}

test_wavs/4.wav
{"start_time":0.00, "keyword": "bo2", "timestamps": [1.12, 1.16], "tokens":["b", "ó"]}

test_wavs/4.wav
{"start_time":0.00, "keyword": "bo2", "timestamps": [1.28, 1.32], "tokens":["b", "ó"]}

[diyism]

And I also tested the "sherpa-onnx-keyword-spotter-microphone", it's amazing, when I say "jiang3 you3 bo2 bei4" to my microphone, it can syllable-level real-time output "jiang3", "you3", "bei4", but missed a syllable of "bo2":

[diyism]

It seems the allosaurus project (https://github.com/xinjli/allosaurus) can produce IPAs and correct timestamps but it missed a syllable of "bei" (in 5 syllables of "jiang3 you3 bo2 bei4 pai1"):

$ python -m allosaurus.run -e 1.2 --lang=cmn --timestamp=True -i ../sherpa-onnx/sherpa-onnx-kws-zipformer-wenetspeech-3.3M-2024-01-01/test_wavs/4_8000hz.wav
0.510 0.045 ɕ
0.540 0.045 i
0.600 0.045 a
0.630 0.045 ŋ

0.750 0.045 j
0.780 0.045 i
0.810 0.045 o
0.870 0.045 ə

1.020 0.045 p
1.080 0.045 o

1.590 0.045 p
1.650 0.045 a
1.710 0.045 i