Top Speaker Computer Vision Models
The models below have been fine-tuned for various speaker detection tasks. You can try out each model in your browser, or test an edge deployment solution (i.e. to an NVIDIA Jetson). You can use the datasets associated with the models below as a starting point for building your own speaker detection model.
At the bottom of this page, we have guides on how to count speakers in images and videos.
Guide: How to Count Speakers with Computer Vision
With a model hosted on Roboflow like the ones above and the open source supervision Python package, you can count speakers in your images and videos.
The following code snippet counts the number of speakers present in an image.
To use the snippet below, you will need to run pip install roboflow supervision. Replace the project name and model name with any model trained on Universe, such as those listed above.
import supervision as sv
import roboflow
roboflow.login()
rf = roboflow.Roboflow()
# replace with the speaker project you choose above
project = rf.workspace("sun-moon-university").project("elevator_button_recognition")
speaker_model = project.version(1).model
results = speaker_model.predict("speaker.jpg").json()
speakers = sv.Detections.from_roboflow(results)
# print number of speakers
print(len(speakers))Guide: How to Count Speakers in a Zone
With a bit more code, you can count the number of speaker present in a specific zone of your image or video.
The following code snippet counts the number of speaker present in each frame in a video.
To use the snippet below, you will need to run pip install roboflow supervision. Replace the project name and model name with any model trained on Universe, such as those listed above.
Read our blog post on counting objects in a zone
import numpy as np
import supervision as sv
import roboflow
SOURCE_VIDEO_PATH = "speaker.mp4"
TARGET_VIDEO_PATH = "speaker_out.mp4"
# use https://roboflow.github.io/polygonzone/ to get the points for your shape
polygon = np.array([
# draw 50x50 box in top left corner
[0, 0],
[50, 0],
[50, 50],
[0, 50]
])
roboflow.login()
rf = roboflow.Roboflow()
# replace with the speaker project you choose above
project = rf.workspace("sun-moon-university").project("elevator_button_recognition")
speaker_model = project.version(1).model
# create BYTETracker instance
speaker_tracker = sv.ByteTrack(track_thresh=0.25, track_buffer=30, match_thresh=0.8, frame_rate=30)
# create VideoInfo instance
video_info = sv.VideoInfo.from_video_path(SOURCE_VIDEO_PATH)
# create frame generator
generator = sv.get_video_frames_generator(SOURCE_VIDEO_PATH)
# create PolygonZone instance
zone = sv.PolygonZone(polygon=polygon, frame_resolution_wh=(video_info.width, video_info.height))
# create box annotator
box_annotator = sv.BoxAnnotator(thickness=4, text_thickness=4, text_scale=2)
colors = sv.ColorPalette.default()
# create instance of BoxAnnotator
zone_annotator = sv.PolygonZoneAnnotator(thickness=4, text_thickness=4, text_scale=2, zone=zone, color=colors.colors[0])
# define call back function to be used in video processing
def callback(frame: np.ndarray, index:int) -> np.ndarray:
# model prediction on single frame and conversion to supervision Detections
results = speaker_model.predict(frame).json()
speakers = sv.Detections.from_roboflow(results)
# show speaker detections in real time
print(speakers)
# tracking speaker detections
speakers = speaker_tracker.update_with_detections(speakers)
annotated_frame = box_annotator.annotate(scene=frame, detections=speakers)
annotated_frame = zone_annotator.annotate(scene=annotated_frame)
# return frame with box and line annotated result
return annotated_frame
# process the whole video
sv.process_video(
source_path = SOURCE_VIDEO_PATH,
target_path = TARGET_VIDEO_PATH,
callback=callback
)Guide: How to Track Speakers Crossing a Line
You can count how many speakers have crossed a line using the supervision LineCounter method.
The following code snippet counts the number of speakers that cross a line in a video.
To use the snippet below, you will need to run pip install roboflow supervision. Replace the project name and model name with any model trained on Universe, such as those listed above.
import numpy as np
import supervision as sv
import roboflow
SOURCE_VIDEO_PATH = "speaker.mp4"
TARGET_VIDEO_PATH = "speaker_out.mp4"
# use https://roboflow.github.io/polygonzone/ to get the points for your line
LINE_START = sv.Point(0, 300)
LINE_END = sv.Point(800, 300)
roboflow.login()
rf = roboflow.Roboflow()
# replace with the speaker project you choose above
project = rf.workspace("sun-moon-university").project("elevator_button_recognition")
speaker_model = project.version(1).model
# create BYTETracker instance
speaker_tracker = sv.ByteTrack(track_thresh=0.25, track_buffer=30, match_thresh=0.8, frame_rate=30)
# create VideoInfo instance
video_info = sv.VideoInfo.from_video_path(SOURCE_VIDEO_PATH)
# create frame generator
generator = sv.get_video_frames_generator(SOURCE_VIDEO_PATH)
# create LineZone instance, it is previously called LineCounter class
line_zone = sv.LineZone(start=LINE_START, end=LINE_END)
# create instance of BoxAnnotator
box_annotator = sv.BoxAnnotator(thickness=4, text_thickness=4, text_scale=2)
# create instance of TraceAnnotator
trace_annotator = sv.TraceAnnotator(thickness=4, trace_length=50)
line_zone_annotator = sv.LineZoneAnnotator(thickness=4, text_thickness=4, text_scale=2)
# define call back function to be used in video processing
def callback(frame: np.ndarray, index:int) -> np.ndarray:
# model prediction on single frame and conversion to supervision Detections
results = speaker_model.predict(frame).json()
speakers = sv.Detections.from_roboflow(results)
# show speaker detections in real time
print(speakers)
# tracking speaker detections
speakers = speaker_tracker.update_with_detections(speakers)
annotated_frame = trace_annotator.annotate(
scene=frame.copy(),
detections=speakers
)
annotated_frame=box_annotator.annotate(
scene=annotated_frame,
detections=speakers
)
# update line counter
line_zone.trigger(speakers)
# return frame with box and line annotated result
return line_zone_annotator.annotate(annotated_frame, line_counter=line_zone)
# process the whole video
sv.process_video(
source_path = SOURCE_VIDEO_PATH,
target_path = TARGET_VIDEO_PATH,
callback=callback
)