Automatic speech recognition (ASR), the conversion of spoken speech to text, is a very important and thriving area of machine learning. ASR algorithms run on practically every smartphone, and are becoming increasingly embedded in professional workflows, such as digital assistants for nurses and doctors. Because ASR algorithms are designed to be used directly by customers and end users, it is important to validate that they are behaving as expected when confronted with a wide variety of speech patterns (different accents, pitches, and background audio conditions).
Using gradio
, you can easily build a demo of your ASR model and share that with a testing team, or test it yourself by speaking through the microphone on your device.
This tutorial will show how to take a pretrained speech-to-text model and deploy it with a Gradio interface. We will start with a full-context model, in which the user speaks the entire audio before the prediction runs. Then we will adapt the demo to make it streaming, meaning that the audio model will convert speech as you speak. The streaming demo that we create will look something like this (try it below or in a new tab!):
Real-time ASR is inherently stateful, meaning that the model’s predictions change depending on what words the user previously spoke. So, in this tutorial, we will also cover how to use state with Gradio demos.
Make sure you have the gradio
Python package already installed. You will also need a pretrained speech recognition model. In this tutorial, we will build demos from 2 ASR libraries:
pip install transformers
and pip install torch
) pip install deepspeech==0.8.2
)Make sure you have at least one of these installed so that you can follow along the tutorial. You will also need ffmpeg
installed on your system, if you do not already have it, to process files from the microphone.
Here’s how to build a real time speech recognition (ASR) app:
First, you will need to have an ASR model that you have either trained yourself or you will need to download a pretrained model. In this tutorial, we will start by using a pretrained ASR model from the Hugging Face model, Wav2Vec2
.
Here is the code to load Wav2Vec2
from Hugging Face transformers
.
from transformers import pipeline
p = pipeline("automatic-speech-recognition")
That’s it! By default, the automatic speech recognition model pipeline loads Facebook’s facebook/wav2vec2-base-960h
model.
We will start by creating a full-context ASR demo, in which the user speaks the full audio before using the ASR model to run inference. This is very easy with Gradio — we simply create a function around the pipeline
object above.
We will use gradio
’s built in Audio
component, configured to take input from the user’s microphone and return a filepath for the recorded audio. The output component will be a plain Textbox
.
import gradio as gr
def transcribe(audio):
text = p(audio)["text"]
return text
gr.Interface(
fn=transcribe,
inputs=gr.Audio(source="microphone", type="filepath"),
outputs="text").launch()
So what’s happening here? The transcribe
function takes a single parameter, audio
, which is a filepath to the audio file that the user has recorded. The pipeline
object expects a filepath and converts it to text, which is returned to the frontend and displayed in a textbox.
Let’s see it in action! (Record a short audio clip and then click submit, or open in a new tab):
Ok great! We’ve built an ASR model that works well for short audio clips. However, if you are recording longer audio clips, you probably want a streaming interface, one that transcribes audio as the user speaks instead of just all-at-once at the end.
The good news is that it’s not too difficult to adapt the demo we just made to make it streaming, using the same Wav2Vec2
model.
The biggest change is that we must now introduce a state
parameter, which holds the audio that has been transcribed so far. This allows us to only the latest chunk of audio and simply append it to the audio we previously transcribed.
When adding state to a Gradio demo, you need to do a total of 3 things:
state
parameter to the functionstate
at the end of the function"state"
components to the inputs
and outputs
in Interface
Here’s what the code looks like:
def transcribe(audio, state=""):
text = p(audio)["text"]
state += text + " "
return state, state
# Set the starting state to an empty string
gr.Interface(
fn=transcribe,
inputs=[
gr.Audio(source="microphone", type="filepath", streaming=True),
"state"
],
outputs=[
"textbox",
"state"
],
live=True).launch()
Notice that we’ve also made one other change, which is that we’ve set live=True
. This keeps the Gradio interface running constantly, so it automatically transcribes audio without the user having to repeatedly hit the submit button.
Let’s see how it does (try below or in a new tab)!
One thing that you may notice is that the transcription quality has dropped since the chunks of audio are so small, they lack the context to properly be transcribed. A “hacky” fix to this is to simply increase the runtime of the transcribe()
function so that longer audio chunks are processed. We can do this by adding a time.sleep()
inside the function, as shown below (we’ll see a proper fix next)
from transformers import pipeline
import gradio as gr
import time
p = pipeline("automatic-speech-recognition")
def transcribe(audio, state=""):
time.sleep(2)
text = p(audio)["text"]
state += text + " "
return state, state
gr.Interface(
fn=transcribe,
inputs=[
gr.Audio(source="microphone", type="filepath", streaming=True),
"state"
],
outputs=[
"textbox",
"state"
],
live=True).launch()
Try the demo below to see the difference (or open in a new tab)!
You’re not restricted to ASR models from the transformers
library — you can use your own models or models from other libraries. The DeepSpeech
library contains models that are specifically designed to handle streaming audio data. These models perform really well with streaming data as they are able to account for previous chunks of audio data when making predictions.
Going through the DeepSpeech library is beyond the scope of this Guide (check out their excellent documentation here), but you can use Gradio very similarly with a DeepSpeech ASR model as with a Transformers ASR model.
Here’s a complete example (on Linux):
First install the DeepSpeech library and download the pretrained models from the terminal:
wget https://github.com/mozilla/DeepSpeech/releases/download/v0.8.2/deepspeech-0.8.2-models.pbmm
wget https://github.com/mozilla/DeepSpeech/releases/download/v0.8.2/deepspeech-0.8.2-models.scorer
apt install libasound2-dev portaudio19-dev libportaudio2 libportaudiocpp0 ffmpeg
pip install deepspeech==0.8.2
Then, create a similar transcribe()
function as before:
from deepspeech import Model
import numpy as np
model_file_path = "deepspeech-0.8.2-models.pbmm"
lm_file_path = "deepspeech-0.8.2-models.scorer"
beam_width = 100
lm_alpha = 0.93
lm_beta = 1.18
model = Model(model_file_path)
model.enableExternalScorer(lm_file_path)
model.setScorerAlphaBeta(lm_alpha, lm_beta)
model.setBeamWidth(beam_width)
def reformat_freq(sr, y):
if sr not in (
48000,
16000,
): # Deepspeech only supports 16k, (we convert 48k -> 16k)
raise ValueError("Unsupported rate", sr)
if sr == 48000:
y = (
((y / max(np.max(y), 1)) * 32767)
.reshape((-1, 3))
.mean(axis=1)
.astype("int16")
)
sr = 16000
return sr, y
def transcribe(speech, stream):
_, y = reformat_freq(*speech)
if stream is None:
stream = model.createStream()
stream.feedAudioContent(y)
text = stream.intermediateDecode()
return text, stream
Then, create a Gradio Interface as before (the only difference being that the return type should be numpy
instead of a filepath
to be compatible with the DeepSpeech models)
import gradio as gr
gr.Interface(
fn=transcribe,
inputs=[
gr.Audio(source="microphone", type="numpy"),
"state"
],
outputs= [
"text",
"state"
],
live=True).launch()
Running all of this should allow you to deploy your realtime ASR model with a nice GUI. Try it out and see how well it works for you.
And you’re done! That’s all the code you need to build a web-based GUI for your ASR model.
Fun tip: you can share your ASR model instantly with others simply by setting share=True
in launch()
.