1 of 21

1

JOHANN

CENTER

RECURSE

DIEDRICK

ZORA

2 of 21

2

Zora

An Interpretable Machine Listening Library for Voice and Speech

Zora

3 of 21

3

AUDREY

Automatic

Digit

Recognizer

Zora

Li, Xiaochang and Mara Mills. "Vocal Features: From Voice Identification to Speech Recognition by Machine." Technology and Culture, vol. 60 no. 2, 2019, p. S129-S160. Project MUSE, https://dx.doi.org/10.1353/tech.2019.0066.

4 of 21

4

MNIST for Speech

Zora

MNIST GIF by Sandipan Dey on SandipanWeb

https://cainvas.ai-tech.systems/use-cases/keras-mnist-vision-app/

0

1

2

7

8

9

3

4

6

5

5 of 21

5

Listener

Zora

6 of 21

6

Listener

Zora

7 of 21

7

Gathering (Data Generation and Pre-Processing)

Zora

8 of 21

8

Gathering (Data Generation and Pre-Processing)

Zora

9 of 21

9

Gathering (Data Generation and Pre-Processing)

Zora

10 of 21

10

Gathering (Data Generation and Pre-Processing)

Zora

11 of 21

11

Learning (Training)

Zora

12 of 21

12

Learning (Training)

Zora

13 of 21

13

Learning (Training)

Zora

14 of 21

14

Listening (Inference)

Zora

15 of 21

15

Listening (Inference)

Zora

16 of 21

16

Listening (Inference)

Zora

17 of 21

17

Understanding (Interpretation)

Zora

18 of 21

18

Understanding (Interpretation)

Zora

19 of 21

19

Understanding (Interpretation)

Zora

20 of 21

20

Listening to Bias

Zora

Li, Xiaochang and Mara Mills. "Vocal Features: From Voice Identification to Speech Recognition by Machine." Technology and Culture, vol. 60 no. 2, 2019, p. S129-S160. Project MUSE, https://dx.doi.org/10.1353/tech.2019.0066.

To close out, I want to return to the Vocal Features paper, where Li and Mills, remark on how these systems were designed to filter our variations in speech for operational efficiencies. Quote:

"The emphasis on phonetic patterns designated Audrey as an extension of ongoing work in speech specification, which sought to isolate the essential “information-bearing elements” that made it possible for humans to understand speech despite pronunciation, pitch, speaking speed, environmental noise, and other forms of acoustic variation.

My hope is that building small, local-first, hyper-specific and personal speech recognition systems can bring about a world in which anyone, no matter their variation in speech performance, can have speech recognition work for them, without the need to modulate one's voice for generalized AI ears.

21 of 21

21

Github: https://github.com/aquietlife/zora

Issues, PRs, and feedback appreciated!

Say hi on Zulip!

@johanndiedrick on IG

Thanks for listening!

Zora