openspeechtointent 0.1.0

A simple, but performant framework for mapping speech directly to categories and intents.

openSpeechToIntent

openSpeechToIntent is a library that maps audio containing speech to pre-specified lists of short texts. It can be used to directly map speech to these texts, which can represent categories or intents for various voice automation applications.

openSpeechToIntent works by using small, but robust speech-to-text models (the default is Citrinet 256 by NVIDIA) to generate predictions for characters or tokens, and then using forced alignment to determine if the audio matches different text transcriptions. The score of the alignment is used to determine which (if any) of the transcriptions is a good match to the audio.

This approach has several useful features:

• No training is required, and any text can be matched to any audio
• It has strong performance while still being relatively fast on a wide range of hardware (see the Performance for more details), as the underlying models are small and efficient
• It can be easily combined with other libraries (e.g., openWakeWord) and tools to create more complex audio processing pipelines

Installation

To install openSpeechToIntent, you can simply use pip:

pip install openSpeechToIntent

This should work on nearly all operating systems (Windows, macOS, Linux), as there are only four requirements: numpy, onnxruntime, pybind11, and sentencepiece.

Usage

openSpeechToIntent is designed to be simple to use. Just provide a file/array of audio data and a list of target intents, and the library will return information about potential intent matches.

from openspeechtointent.model import CitrinetModel

# Load model (this will also download the model if it is not already present)
mdl = CitrinetModel()

# Define some simple intents
intents = ["turn on the light", "pause the music", "set a 5 minute timer"]

# Load a sample audio file (from the test/resources directory in this repo)
# Can also directly provide a numpy array of 16-bit PCM audio data
audio_file = "test/resources/sample_1.wav"  # contains the speech "turn on the lights"

# Match the audio to the provided intents
matched_intents, scores, durations = mdl.match_intents(audio_file, intents)

# View the results
for intent, score, duration in zip(matched_intents, scores, durations):
    print(f"Intent: {intent}, Score: {score}, Duration: {duration}")

# Output:
# Intent: "turn on the lights", Score: 0.578, Duration: 0.800 seconds
# Intent: "pause the music", Score: 0.270, Duration: 0.560 seconds
# Intent: "set a 5 minute timer", Score: 0.119, Duration: 0.560 seconds
# Intent: "remind me to buy apples tomorrow", Score: 0.032, Duration: 1.840 seconds

Scores are computed from the softmaxed logits from the Citrinet model, to scale them between 0 and 1. The score can be used to select possible matching intents (or none at all) by appropriate thresholds.

The durations are the approximate length of the intent as aligned to the audio. This can provide another way to filter and select possible intent matches by selecting those that have the most appropriate duration.

Performance

For many use-cases, the performance of openSpeechToIntent can be surprisingly good. This is a testament to both the high quality of Nvidia pre-trained models, and the way that constraining the speech-to-text decoding to a fixed set of intents greatly reduces the search space of the problem. While real-world performance numbers always depend on the deployment environment, here are some examples use-cases that illustrate the type of performance possible with openSpeechToIntent:

Ignoring false wake word activations

Wake word detection frameworks like openWakeWord, microWakeWord, etc. are designed to efficiently listen for target activation words while continuously processing input audio. The challenge with these types of systems is maintaining high recall of the target activation words, while not activating on other, unrelated audio. In practice, this is a difficult balance that requires careful training and tuning, and performance can vary widely depending on the environment and the specific wake word.

One approach to improving the effective performance of these types of systems is to tune the wakeword model to be very sensitive, and then filter out any false activations through other means. openSpeechToIntent can be used in this way, assuming that there is a known list of intents that would normally be expected after a wake word activation. As an example, the table below shows the performance of the pre-trained Alexa wake word model from openWakeWord on the 24 hour PicoVoice wake word benchmark dataset, where the model's threshold is set very low (0.1) to ensure very high recall but low precision. With this configuration, false positive rate on the Picovoice dataset is ~3.58, unnacceptably high. However, by using openSpeechToIntent to verify that the speech after the activation matches a list ~400 expected intents (see the list here), the false positive rate can be reduced to <0.04 false activations per hour.

openWakeWord Model	openWakeWord Score Threshold	openSpeechToIntent Score Threshold	False Positives per Hour
Alexa	0.1	NA	~3.58
Alexa	0.1	0.1	<0.04

Precisely matching intents

openSpeechToIntent can also be used to perform more fine-grained classification across a reasonable number of intents. As an example, on the Fluent Speech Commands test set, an average accuracy of ~98% across 31 intents is possible.

Model	Accuracy
openSpeechToIntent	98.2%
SOTA	99.7%

Efficiency

openSpeechToIntent is designed to be reasonably efficient, and can run on a wide range of hardware included normal desktop CPUs and moderately powerfull SBCs. The table below shows the performance of the default Nvidia Citrinet model on a several different systems, using a 4 second audio clip as input and matching against 400 intents.

CPU	Number of Threads	Time to Process 4s Audio Clip (ms)	Time to Match Against 400 Intents (ms)
Intel Xeon W-2123	1	103	21
AMD Ryzen 1600	1	98	17
Raspberry Pi 4	1	320	56
Raspberry Pi 4	2	262	56

Note that further optimizations are possible, and in general the length of the audio clip and the number of intents will have the largest impact on the efficiency of the system. By using hueristics to limit the number of intents to match and precisely controlling the length of the audio clip, performance can be further improved.

Advanced Usage

Using raw logit scores

In some cases, instead of applying the softmax transform to the scores, it may be useful to use the raw logit scores directly. For example, if you are primarily using openSpeechToIntent to filter out false wake word activations from another system, the raw logit score can make it easer to set a global threshold that works well across all intents.

As an example, suppose you have a set of 5 intents with raw logits scores of [-6, -7, -8, -9, -10]. In absolute terms, these scores are quite low, and none of the intents have good alignments with the audio. However, applying softmax to these scores gives [0.6364, 0.2341, 0.0861, 0.0317, 0.0117], which falsely implies that the first intent is a good match. By carefully setting a threshold on the raw logit scores by tuning against your specific use-case and deployment environment, you can often achieve better performance compared to using the softmax scores.

Limitations

Currently, the library only supports matching English speech to english intents. Future work may involve expanding to other languages and supporting other speech-to-text frameworks like whisper.cpp.

Acknowledgements

Many thanks to Nvidia for the excellent Citrinet speech-to-text models, as well as many other highly performant speech and audio models.

Also, credit to @MahmoudAshraf97 for the excellent modification of the torch forced alignment cpp functions to simplify dependencies and enable easy usage with pybind.

License

This code in this project is licensed under the Apache 2.0 License - see the LICENSE file for details. Portions of the code adapted in whole or part from other repositories are licensed under their respective licenses, as appropriate.

The Nvidia Citrinit models is licensed under the CC-BY-4.0 license and the NGC Terms of Use.