Build your own real-time voice command recognition model with TensorFlow

[@donahue1187]

This is fantastic. I’m a Newbie to Python and neural nets, but your explanations are great and pretty straightforward. Question - what additional steps would I take to run this on my own local device (pi 4)? And what else would I need to do to introduce new commands such as as trigger word and “turn off the lights”? Would I need to create my own audio samples, save them to new folders, and retrain to retrain the model? Thanks for any guidance! (if you couldn’t tell I’m DONE w Google Home latency, recreating my own. Ambitious! Need help!)

[@cornpop3340]

This is an incredibly helpful video.

[@nguyent3465]

The code on TensorFlow website was changed :(

[@seanadin386]

Can you do a video regarding the newer version? The run interface now has a different code

[@tvartalk]

😊

[@gokhanersoz5239]

Thank you very much for the tranings. But I think there should be a more complex and more advanced voice recognition, voice classification and similar training series if you see fit. You know, trainings on sound are limited.

[@Cyka_Blyatus]

What did you do so the program does not picks up ambient noise or actually works with the commands given? it seems the model lacks ambient noise data sets and whenever ran it only keeps spamming the first command, but yours works perfectly, how to achieve this?

[@obi666]

I have the same issue, files from datasets works perfectly but when I try to use my mic like in the video or record audio with my mic using pyaudio it gives me the 1st class all the time.

[@erickd4816]

Good video, excellent explanation, I have a question, can the same program be trained to recognize only a specific voice? if so, could you explain it to me? I would be very grateful.

[@clumsycoder1907]

its not working for me

[@geekyprogrammer4831]

Can you please post building text to speech models from scratch?

[@MrIlvis]

On which Tensorflow version this was made? because Colab uses latest, but older one should work without problems.

[@obi666]

Colab has unix like OS, you can use commands like pip install, dnf install and etc

[@loydvincentbutron4345]

is it for english voice only?

[@swasthikk3655]

Can i get similar for English alphabets

[@danielbogemann1598]

They changed the Code. Could u you do a quick update?

[@TheSaukkio]

How can it be that in the video it gives nothing with out speaking. While if i run the code from github it predicts random stuff when im not speaking.

[@obi666]

I have the same problem - works perfectly when using audio files from datasets but doesn't when I make some input using my mic

[@obi666]

I've managed to fix that problem - here are the fragments of my code: import sounddevice as sd @staticmethod def record_audio(filename: Union[str, None] = None, duration: int = Config.DURATION, fs: int = Config.INPUT_LEN) -> np.array: print(f"Record sound for {duration} seconds...") audio = sd.rec(int(duration * fs), samplerate=fs, channels=1, dtype=np.float32) sd.wait() #sd.play(audio, samplerate=fs) #sd.wait() # print(f"Recording has been completed") # Convert audio to 16-bit PCM format (in the range [-32768, 32767]) audio_pcm = np.int16(audio * 32767) if filename: write(filename, fs, audio_pcm) # print(f"Sounds has been saved to: {filename}") return np.array([]) return audio_pcm def preprocess_mic_data(self, waveform: np.array) -> EagerTensor: waveform = tf.squeeze(waveform, axis=-1) spectrogram = self.get_spectrogram(waveform=waveform) normalized_spectrogram = tf.expand_dims(spectrogram, axis=0) return normalized_spectrogram def get_prediction(self, audio_data: Union[str, os.PathLike, np.array]) -> str: if isinstance(audio_data, str) or isinstance(audio_data, os.PathLike): normalized_spectrogram = self.preprocess_file_data(file_path=audio_data) else: normalized_spectrogram = self.preprocess_mic_data(waveform=audio_data) prediction = self.model(normalized_spectrogram) # print(prediction) classid = np.argmax(prediction) return self.commands[classid] @staticmethod def calculate_rms(signal: np.ndarray) -> float: """Oblicza RMS (Root Mean Square) sygnału.""" return np.sqrt(np.mean(signal ** 2)) def is_voice_present(self, signal: np.ndarray, rms_threshold: float = 20) -> bool: """ Sprawdza, czy w sygnale znajduje się głos, na podstawie RMS. :param signal: Tablica NumPy z nagranym dźwiękiem. :param rms_threshold: Próg RMS, poniżej którego uznaje się, że nie ma mowy. :return: True, jeśli wykryto głos; False w przeciwnym razie. """ rms_value = self.calculate_rms(signal) print(f"Wartość RMS: {rms_value}") return rms_value > rms_threshold # Sprawdź, czy RMS jest większe niż próg while True: res = predictor.record_audio() print(res.shape) voice = predictor.is_voice_present(signal=res) if voice: print(predictor.get_prediction(audio_data=res)) it's not perfect but at least it works

[@TheSaukkio]

@@obi666 can you send me your code somehow🤔 or contact me so i can get this🤣🤣

@@obi666 masz moze jeszcze to rozwiazanie? bo widze ze z polski jestes

[@obi666]

Dałem odp, nwm czy cię oznaczylo

[@oxydol3456]

This tutorial is great. I find that the key to build accurate model is gathering quality data a lot. And that sounds arduous work. didn't get good result with 200 examples. Edit: I found the model's accuracy is the way poor than I expected. Maybe it's due to the microphone I'm using and it's needed to taken care of before predicting process.

[@tankado_ndakota]

Got the error: "Could not import the PyAudio C module 'pyaudio._portaudio'." And couldn't find the solution... Macbook M1 Pro

[@tankado_ndakota]

I saw a note in other video for M1 :) let me try first :D

[@tankado_ndakota]

i did everything that I found from web. but still i got the error: "symbol not found in flat namespace '_PaMacCore_SetupChannelMap'"

[@sanjeetjha9177]

Please provide me the model i need argently I am stuck in it

[@rediet.f261]

what is sample_file in here 8:38

[@clumsycoder1907]

same doubt

[@arqamrafay]

exactly, i think their is file of recorded audio

[@LukasKofler]

See the first line at 5:38 🙂

[@itsrairamones]

thankyou dude its a hundred percent work for me but after couple minutes it crashed :(

[@Yvtq8K3n]

Its a shame, you cant train your own model.

[@threepe0]

of course you can

[@Yvtq8K3n]

@@threepe0 The last time I used this, you were unable to create a custom model and use it. Tensorflow provided you with an already trained model (0-1, left, right) and thats exactly what most people use.