Example on README doesn't work

[abhinavkulkarni]

Hi,

The following example on top level README doesn't work:

from sources import MicrophoneAudioSource
from functional import FrameWiseModel, ChunkWiseModel, OverlappedSpeechPenalty, EmbeddingNormalization

mic = MicrophoneAudioSource(sample_rate=16000)

# Initialize independent modules
segmentation = FrameWiseModel("pyannote/segmentation")
embedding = ChunkWiseModel("pyannote/embedding")
osp = OverlappedSpeechPenalty(gamma=3, beta=10)
normalization = EmbeddingNormalization(norm=1)

# Branch the microphone stream to calculate segmentation
segmentation_stream = mic.stream.pipe(ops.map(segmentation))
# Join audio and segmentation stream to calculate speaker embeddings
embedding_stream = rx.zip(mic.stream, segmentation_stream).pipe(
    ops.starmap(lambda wave, seg: (wave, osp(seg))),
    ops.starmap(embedding),
    ops.map(normalization)
)

embedding_stream.suscribe(on_next=lambda emb: print(emb.shape))

mic.read()

First of all there is a type in declaring variable ops (it's been declared as osp).

Secondly, in line segmentation_stream = mic.stream.pipe(ops.map(segmentation)), there is no method named map in variable ops.

Thanks!

[KannebTo]

I came accross the same problem. ops is not osp. It is rx.operators. Here are the missing imports:

import rx
import rx.operators as ops

And there is a typo: suscribe should be subscribe.

[abhinavkulkarni]

Hey @KannebTo:

Thanks. I renamed osp go ops. However, the other error with related there being no method named map in variable ops still persists.

[juanmc2005]

Hello, thank you for reporting these typos.
As mentioned by @KannebTo, ops references rx.operators and it's indeed subscribe.

As for the map method error, I'll try to reproduce it and get back to you.

[KannebTo]

Hey @abhinavkulkarni , just leave osp and ops as in the original code and add the imports on the top.

[abhinavkulkarni]

Hey @KannebTo,

Here's the error I get after renaming osp to ops and adding the import statements at the top:

>>> import rx
>>> import rx.operators as ops
>>> from sources import MicrophoneAudioSource

/home/user/miniconda3/envs/diarization/lib/python3.8/site-packages/torchaudio/backend/utils.py:46: UserWarning: "torchaudio.USE_SOUNDFILE_LEGACY_INTERFACE" flag is deprecated and will be removed in 0.9.0. Please remove the use of flag.
  warnings.warn(
>>> from functional import FrameWiseModel, ChunkWiseModel, OverlappedSpeechPenalty, EmbeddingNormalization
>>> 
>>> mic = MicrophoneAudioSource(sample_rate=16000)
>>> 
>>> # Initialize independent modules
... segmentation = FrameWiseModel("pyannote/segmentation")
>>> embedding = ChunkWiseModel("pyannote/embedding")
>>> ops = OverlappedSpeechPenalty(gamma=3, beta=10)
>>> normalization = EmbeddingNormalization(norm=1)
>>> 
>>> # Branch the microphone stream to calculate segmentation
... segmentation_stream = mic.stream.pipe(ops.map(segmentation))
Traceback (most recent call last):
  File "<stdin>", line 2, in <module>
AttributeError: 'OverlappedSpeechPenalty' object has no attribute 'map'
>>> # Join audio and segmentation stream to calculate speaker embeddings
... embedding_stream = rx.zip(mic.stream, segmentation_stream).pipe(
...     ops.starmap(lambda wave, seg: (wave, osp(seg))),
...     ops.starmap(embedding),
...     ops.map(normalization)
... )
Traceback (most recent call last):
  File "<stdin>", line 2, in <module>
NameError: name 'segmentation_stream' is not defined
>>> 
>>> embedding_stream.suscribe(on_next=lambda emb: print(emb.shape))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
NameError: name 'embedding_stream' is not defined
>>> 
>>> mic.read()

I am running this from the src directory.

[KannebTo]

ops = OverlappedSpeechPenalty(gamma=3, beta=10) must be osp as in the original

[KannebTo]

from sources import MicrophoneAudioSource
from functional import FrameWiseModel, ChunkWiseModel, OverlappedSpeechPenalty, EmbeddingNormalization
import rx
import rx.operators as ops

mic = MicrophoneAudioSource(sample_rate=16000)

# Initialize independent modules
segmentation = FrameWiseModel("pyannote/segmentation")
embedding = ChunkWiseModel("pyannote/embedding")
osp = OverlappedSpeechPenalty(gamma=3, beta=10)
normalization = EmbeddingNormalization(norm=1)

# Branch the microphone stream to calculate segmentation
segmentation_stream = mic.stream.pipe(ops.map(segmentation))
# Join audio and segmentation stream to calculate speaker embeddings
embedding_stream = rx.zip(mic.stream, segmentation_stream).pipe(
    ops.starmap(lambda wave, seg: (wave, osp(seg))),
    ops.starmap(embedding),
    ops.map(normalization)
)

embedding_stream.subscribe(on_next=lambda emb: print(emb.shape))

mic.read()

[juanmc2005]

This should be fixed now.
Please let me know if it works from your end.

[abhinavkulkarni]

Thanks, it works!

[KannebTo]

Thanks! Yes it works with the regularize_stream().

But i have another issue with the audio access, probably due to my audio setup:

>>> mic.read()
torch.Size([4, 512])
python: src/hostapi/alsa/pa_linux_alsa.c:3641: PaAlsaStreamComponent_BeginPolling: Assertion `ret == self->nfds' failed.
Aborted (core dumped)

This is strange because I'm running this in a Docker container and I also use the sounddevice library without problems in another project in the same container.

[juanmc2005]

@KannebTo would you mind creating a separate issue for this? Thanks!