Using TeachableMachine Trained Model in python after converting to .h5 and .pb files

Hey everyone, my first post here…

i made a wake word detecting model at teachable machine and it works perfectly as a TensorflowJs model but when I convert it to a .tflite the accuracy goes down (basically stops working) so I converted the tensorflowjs model into a tensorflow saved model using the tensorflow_convertor in WSL. then this tensorflow saved model was converted to a hd5 model, and now I have no idea how to use this model offline.

import time
import numpy as np
import pyaudio
import tensorflow as tf

# Load the saved model
loaded_model = tf.saved_model.load('./keras')

# Get the serving default signature
serving_default = loaded_model.signatures["serving_default"]

# Define parameters for microphone input
FORMAT = pyaudio.paInt16  # Sample format
CHANNELS = 1  # Number of audio channels (1 for mono, 2 for stereo)
RATE = 16000  # Sampling rate (samples per second)
CHUNK = 1024  # Number of frames per buffer
DESIRED_DURATION_SECONDS = 0.6  # Desired duration of audio for each prediction

# Create an instance of PyAudio
audio = pyaudio.PyAudio()

# Open the microphone stream
stream =,


# Start timing
start_time = time.time()

# Buffer to hold microphone input
frames = []

# Function to preprocess audio data
def preprocess_audio(audio_data, target_shape=(43, 232)):
    Preprocesses audio data to match the input shape expected by the model.

    - audio_data: Raw audio data as numpy array.
    - target_shape: Target shape of the audio data after preprocessing.

    - Preprocessed audio data as numpy array.
    # Perform any necessary preprocessing steps (e.g., normalization)
    normalized_audio = audio_data / np.iinfo(np.int16).max  # Normalize to range [-1, 1]

    # Determine the number of frames per row
    frames_per_row = target_shape[1]

    # Calculate the number of rows needed to accommodate all frames
    num_rows = int(np.ceil(len(normalized_audio) / frames_per_row))

    # Pad the audio data if necessary to ensure it fits into the target shape
    padded_audio = np.pad(normalized_audio, (0, num_rows * frames_per_row - len(normalized_audio)), mode='constant')

    # Reshape the padded audio data to match the target shape
    preprocessed_audio = np.reshape(padded_audio, (num_rows, frames_per_row))

    # Add batch and channel dimensions
    preprocessed_audio = np.expand_dims(preprocessed_audio, axis=0)  # Add batch dimension
    preprocessed_audio = np.expand_dims(preprocessed_audio, axis=-1)  # Add channel dimension

    return preprocessed_audio

# Record audio data in chunks and make predictions
    while True:
        # Read microphone data
        data =

        # When enough data is collected (adjust CHUNK size for desired duration)
        if len(frames) == int(DESIRED_DURATION_SECONDS * RATE / CHUNK):
            # Convert microphone data to numpy array
            audio_data = np.frombuffer(b''.join(frames), dtype=np.int16)

            # Preprocess audio data
            preprocessed_audio = preprocess_audio(audio_data)

            # Make predictions
            predictions = serving_default(conv2d_1_input=tf.constant(preprocessed_audio, dtype=tf.float32))

            # Extract probabilities
            background_noise_prob = predictions['sequential_3'].numpy()[0][0]
            wake_word_prob = predictions['sequential_3'].numpy()[0][1]

            # Print results
            print("Background Noise Probability:", round(background_noise_prob * 100), "%")
            print("Wake Word Probability:", round(wake_word_prob * 100), "%")

            # End timing
            end_time = time.time()

            # Calculate elapsed time
            elapsed_time_ms = (end_time - start_time) * 1000

            # Print elapsed time
            print("Time Taken:", round(elapsed_time_ms, 1), "ms.")

            # Reset frames for the next iteration
            frames = []

except KeyboardInterrupt:
    # Stop the stream when Ctrl+C is pressed
    print("Stopped recording.")

# Close the stream and PyAudio

Edit: I managed to get the saved model (.pb) in python with gpt’s help but its not working properly.

Modified by moderator

AFAIK we do not support conversion from TensorFlow.js to TF? Did you use a 3p tool for doing that?

I have referred to tis link: “How to convert tensorflow.js model and weights to standard tensorflow? - Stack Overflow

Used the following command: tensorflowjs_converter --input_format=tfjs_layers_model --output_format=keras path/to/tfjs_files path/to/save_keras_hdf5

I did not need to install any libraries apart from tensorflowjs and tensorflow so ig its not 3rd party right?

For some reason I thought we did not support going that way around.

@pyu Will have the definitive answer though who worked on the converter.

1 Like

Thank you @Jason… I hope this will be resolved soon…

Any updates on this @Jason?

This should still work, but it will create keras v2 model.

Can I please get some sample code on how to provide the audio data from the microphone to the model and get the outputs? I’ve been trying but none of the codes seem to be working…

So ig i have to train a model manually using tensorflow python and have to scrap this… ohk