In our previous article, we discussed how to unlock some of that data by building a speaker diarization system for online meetings (POC) to identify speakers in audio streams and provide organizations with detailed speaker-based insights into meetings, create meeting summaries, action items, and more.
However, effective communication goes beyond words. A subtle non-verbal cue or underlying emotional state can reveal how something was said and entirely change the meaning.
Analyzing emotions from various voice interactions such as customer service calls, sales meetings, or online interviews can help unlock deeper insights to predict behavior, take data-driven actions based on expected behavior, and improve quality monitoring over time.
For services like contact centers or sales-focused meeting platforms, this can turn into increased sales reps' performance, personalized customer assistance, better understanding of customer satisfaction, and more.
In this tutorial, we give you building blocks for integrating advanced speaker-based summaries for online meetings using speaker diarization and emotion analysis. Let’s dive in.
Advanced speaker diarization identifies and segments an audio recording into speech segments, where each segment corresponds to a specific speaker. It then detects changes in speaker identity and groups segments that belong to the same speaker, answering the question who spoke when.
Emotion analysis, as the name suggests, analyzes the emotional undertones of voice, and classifies them into categories such as approval, disappointment, excitement, and curiosity, answering the question how was that said. For emotion analysis in this tutorial, we’ll use a version of Whisper called Whisper-timestamped and Hugging Face emotion detection model.
Sentiment vs emotion analysis
Note: Emotion analysis is often confused with sentiment analysis. Sentiment analysis classifies information as positive, negative, or neutral. But it’s not always capable of identifying emotional nuances such as surprise or fear. Here's where emotion analysis comes in—it analyzes more complex emotions and undertones. Both sentiment and emotion analysis can be text or speech-based.
This POC has various use cases:
Here are the steps for implementing advanced speaker diarization and emotion analysis POC:
Ready? Let’s get started.
Setting up a Python virtual environment helps you manage dependencies and keep your projects organized. You’ll need Python 3.6 or later:
pip install virtualenvCreate virtual environment virtualenv env
virtualenv envActivate virtual environment
Windows:
env\Scripts\activatemacOS and Linux:
source env/bin/activateOptional: While optional, Google Colab is recommended for running notebooks in an environment with a free GPU. This is especially useful for processing large audio files efficiently.
We’ll use several libraries and packages in this tutorial.
Pyannote.audio: Speaker diarization with pre-trained models for segmenting and labeling speaker identities. Install it using pip:
!pip install pyannote.audioSpeechBrain: All-in-one, open-source speech toolkit enabling flexible speech technology experiments. Install it directly from the GitHub repository:
!pip install git+https://github.com/speechbrain/speechbrain.git@developTorchaudio: Gives you access to audio files and transformations:
!pip install torchaudioSciPy: Used for scientific computing; it helps in operations like computing distances between embeddings:
!pip install scipyHugging Face Transformers: A wealth of pre-trained models used for ASR and emotion analysis tasks:
!pip install transformersWhisper-timestamped: An enhanced ASR model that offers precise transcription with timestamps, critical for synchronizing transcribed text with audio segments:
!pip install whisper-timestampedEnsure you have audio recordings for analysis. You can use the audio library featured in this tutorial if you don't have your recordings.
Let’s start by creating unique speaker embeddings for known speakers:
import torch
import torchaudio
from speechbrain.inference.speaker import EncoderClassifier
from pyannote.audio import Pipeline, Audio
from scipy.spatial.distance import cdist
# Check if CUDA is available and set the device accordingly
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Load pre-trained model for speaker embedding extraction and move it to the device
# Note: You need to obtain an API key from Hugging Face to use this model.
classifier = EncoderClassifier.from_hparams(source="speechbrain/spkrec-ecapa-voxceleb", run_opts={"device": device})
classifier = classifier.to(device)
# Pre-trained model for speaker diarization
# Note: The speaker diarization model also requires an API key from Hugging Face.
# YOUR_HUGGING_FACE_API_KEY
diarization = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1",
use_auth_token="Your_API_Key")Now we need to:
# Extract embeddings for known speakers and move them to the device
known_speakers = []
known_speaker_ids = [] # To keep track of speaker IDs
for speaker_id, speaker_file in enumerate(["/content/Steve Job's sample.wav", "/content/elon musk sample.wav", "/content/mandela sample.wav"]):
waveform, sample_rate = torchaudio.load(speaker_file)
waveform = waveform.to(device)
embedding = classifier.encode_batch(waveform)
known_speakers.append(embedding.squeeze(1).cpu().numpy()) # Squeeze and move to CPU
# Update speaker labels here
if speaker_id == 0:
known_speaker_ids.append("Steve Jobs")
elif speaker_id == 1:
known_speaker_ids.append("Elon Musk")
elif speaker_id == 2:
known_speaker_ids.append("Nelson Mandela")Segment the audio file into different speaker segments to identify "who spoke when."
# Process meeting recording
segments = diarization("/content/mandela elon steve.wav")Next we need to identify speakers in each segment of the recording. For each segment, we load the corresponding portion of the audio recording using torchaudio.load() and extract the waveform.
We then pass the waveform to the classifier model to obtain the speaker embedding using the encode_batch() method. The obtained embedding is compared with embeddings of known speakers using cosine distance.
The speaker with the minimum distance is identified as the speaker for the specific segment:
# Set the threshold for similarity scores
threshold = 0.8
# Function to process each segment and identify speakers
def identify_speakers_in_segments(segments, known_speakers, known_speaker_ids, classifier, threshold):
identified_segments = []
for i, (segment, label, confidence) in enumerate(segments.itertracks(yield_label=True)):
start_time, end_time = segment.start, segment.end
waveform, sample_rate = torchaudio.load("/content/mandela elon steve.wav")
start_index = int(start_time * sample_rate)
end_index = int(end_time * sample_rate)
waveform_segment = waveform[:, start_index:end_index].to(device)
# Ensure the waveform segment is appropriately padded if necessary
if waveform_segment.size(1) < 16000: # Assuming 1 second as minimum length
pad_size = 16000 - waveform_segment.size(1)
waveform_segment = torch.nn.functional.pad(waveform_segment, (0, pad_size))
embedding = classifier.encode_batch(waveform_segment)
embedding_2d = embedding.squeeze(1).cpu().numpy()
min_distance = float('inf')
recognized_speaker_id = None
# Iterate over each known speaker and find the minimum distance for j, speaker_embedding in enumerate(known_speakers):
distances = cdist(embedding_2d, speaker_embedding, metric="cosine") # No need to reshape for cdist
min_distance_candidate = distances.min()
if min_distance_candidate < min_distance:
min_distance = min_distance_candidate
recognized_speaker_id = known_speaker_ids[j] # Get the corresponding speaker ID
if min_distance < threshold:
segment_info = {
"segment_id": i,
"speaker": recognized_speaker_id,
"begin_time": start_time,
"end_time": end_time
}
identified_segments.append(segment_info)
else:
segment_info = {
"segment_id": i,
"speaker": "Unknown",
"begin_time": start_time,
"end_time": end_time
}
identified_segments.append(segment_info)
return identified_segments
# Call the function with appropriate parameters
identified_segments=identify_speakers_in_segments(segments, known_speakers, known_speaker_ids, classifier, threshold)Whisper-timestamped is based on OpenAI Whisper. However, it predicts word timestamps and provides a more accurate estimation of speech segments compared to Whisper models.
A confidence score is assigned to each word and each segment.
import whisper_timestamped as whisper
audio = whisper.load_audio("/content/mandela elon steve.wav")
model = whisper.load_model("large-v3")
result = whisper.transcribe(model, audio)Use the following snippet codes to assign the transcription segment to a right speaker.
# Function to process the ASR output
def process_asr_output(asr_data):
return asr_data['segments']
# Function to process the diarization and speaker identification output
def process_speaker_segments(speaker_data):
return speaker_data
# Function to assign speakers to transcription segments
def assign_speakers_to_transcriptions(trans_segments, speaker_segments):
for trans in trans_segments:
trans_start = trans['start']
trans_end = trans['end']
trans['speaker'] = None
for speaker in speaker_segments:
if speaker['begin_time'] <= trans_start < speaker['end_time'] or \
speaker['begin_time'] < trans_end <= speaker['end_time']:
trans['speaker'] = speaker['speaker']
break
return trans_segments
# Main processing steps
transcription_segments = process_asr_output(result)
speaker_segments = process_speaker_segments(identified_segments)
updated_transcriptions = assign_speakers_to_transcriptions(transcription_segments, speaker_segments)
# Prepare data for JSON output
transcription_w_speakers_data = []for trans in updated_transcriptions:
transcription_w_speakers_data.append({
"id": trans['id'],
"timestamp": {"start": trans['start'], "end": trans['end']},
"speaker": trans['speaker'],
"text": trans['text']
})
# Convert list to JSON string
json_string = json.dumps(transcription_w_speakers_data, indent=4)
# Print JSON string
print(json_string)
We’ll use a pre-trained model, "SamLowe/roberta-base-go_emotions", from the Hugging Face Transformers library. This model is trained to recognize a wide range of emotions from text inputs.
from transformers import pipeline
sentiment_analysis = pipeline(
"sentiment-analysis",
framework="pt",
model="SamLowe/roberta-base-go_emotions"
)
def analyze_sentiment(text):
results = sentiment_analysis(text)
sentiment_results = {
result['label']: result['score'] for result in results
}
return sentiment_resultsThe analyze_and_append_sentiments function goes through transcription data, analyzes the sentiment of each segment, and assigns the results to the corresponding segments in the JSON data:
# Function to analyze sentiment and update the JSON data
def analyze_and_append_sentiments(transcription_w_speakers_data):
for segment in transcription_w_speakers_data:
text = segment['text']
sentiment_results = analyze_sentiment(text)
segment['sentiment'] = sentiment_results # Append sentiment result to each segment
return transcription_w_speakers_data# Analyze sentiments and update JSON data
updated_json_data = analyze_and_append_sentiments(transcription_w_speakers_data)
# Print updated JSON data with sentiment results
json_string = json.dumps(updated_json_data, indent=4)
print(json_string)Here is the final result:
[
{
"id": 0,
"timestamp": {
"start": 0.0,
"end": 3.06
},
"speaker": "Nelson Mandela",
"text": " Sport has the power to change the world.",
"sentiment": {
"optimism": 0.5600042343139648
}
},
{
"id": 1,
"timestamp": {
"start": 3.66,
"end": 4.88
},
"speaker": null,
"text": " It has the power...",
"sentiment": {
"neutral": 0.8447567820549011
}
},
{
"id": 2,
"timestamp": {
"start": 5.3,
"end": 6.32
},
"speaker": "Elon Musk",
"text": " Which I get asked a lot.",
"sentiment": {
"neutral": 0.9046566486358643
}
},
{
"id": 3,
"timestamp": {
"start": 6.98,
"end": 7.48
},
"speaker": "Nelson Mandela",
"text": " ...to inspire.",
"sentiment": {
"neutral": 0.9702799916267395
}
},
{
"id": 4,
"timestamp": {
"start": 8.46,
"end": 14.66
},
"speaker": "Nelson Mandela",
"text": " It has the power to unite people in a way that little else does.",
"sentiment": {
"approval": 0.528308093547821
}
},
{
"id": 5,
"timestamp": {
"start": 15.56,
"end": 18.54
},
"speaker": "Steve Jobs",
"text": " It's a breakthrough internet communications device.",
"sentiment": {
"neutral": 0.4549199938774109
}
},
{
"id": 6,
"timestamp": {
"start": 20.06,
"end": 24.36
},
"speaker": "Nelson Mandela",
"text": " It speaks to youth in a language they understand.",
"sentiment": {
"neutral": 0.8397235870361328
}
},
{
"id": 7,
"timestamp": {
"start": 25.0,
"end": 27.2
},
"speaker": "Elon Musk",
"text": " When you think about the question of where are the aliens,",
"sentiment": {
"neutral": 0.7901335954666138
}
}
]
The advanced speaker diarization and emotion analysis POC enhances the processing and analyzing of online meetings, making them more accessible, organized, and efficient.
However, this model has its limitations. It currently processes transcribed text without considering the rich audio cues such as tone, pitch, and energy that can alter the emotion of a spoken sentence — the same sentence spoken in a cheerful tone versus an angry tone can convey entirely different emotions.
In our next tutorial, we will show you how to overcome this limitation. You’ll learn how to integrate models that can analyze emotions directly from audio data and leverage auditory cues to provide a more nuanced and accurate analysis. Stay tuned!
Gladia provides a speech-to-text and audio intelligence API for building virtual meeting and note-taking apps, call center platforms, and media products, providing transcription, translation, and insights powered by best-in-class ASR, LLMs and GenAI models.
Speech-To-Text
Over the last few years, Large Language Models (LLMs) have become accessible and transformative tools, powering everything from customer support and content generation to complex, industry-specific applications in healthcare, education, and finance.
Case Studies
In recent years, fuelled by advancements in LLMs, the numbers of AI note-takers has skyrocketed. These apps are increasingly tailored to meet the unique needs of specific user groups, such as doctors, sales teams and project managers.
Speech-To-Text
It happens all the time. A user submits a query to a large language model (LLM) and swiftly gets a response that is clear, comprehensive, and obviously incorrect.
