Meeting Transcription Deep Dive: How It Works, What's Coming Next

Companies are recording their calls and meetings more than ever before. These recordings contain critical insights into the thoughts, feelings, and opinions of prospects and customers. However, finding these insights can be difficult, especially as the volume of calls recorded continues to grow. Manually reviewing calls is time-consuming, expensive, and does not scale. Meeting transcription helps ease the burden of extracting actionable value from call recordings by transforming the raw audio into searchable text that can subsequently be analyzed.

Transcribing calls and meetings requires a sophisticated combination of different AI/ML technologies, including automatic speech recognition (ASR), with many other vital components. This article provides an overview of how meeting transcription works, the key technologies involved, and how the conversational intelligence industry is moving beyond basic meeting transcription. 

Meeting transcription is the process of converting audio into text. Meetings involve rapid-fire conversations between multiple individuals who may be in the same room or dialed in remotely. These conversations include video and screen sharing that can be very information-dense. A suitable meeting transcript involves more than just understanding the words that were said: It must preserve the flow of the conversation, recognize critical events that occurred, and understand the intent of what was said. Good transcripts include speaker naming and statements to understand who was involved in discussion points while giving specific context to what was said.

Quality meeting transcription systems are built with many required components, including simple acoustic speech recognition (ASR). These systems must provide conversation diarization and speaker identification support to automatically identify the names of new meeting participants for proper attribution. They must provide natural language understanding to split words into sentences or paragraphs, capitalize and understand acronyms, and clean up the transcribed text. Verbal communication contains many unique challenges not found in text-based communication, such as disfluencies, stuttering, word repetitions, background noise, vocal cues, and non-verbal communication. A quality meeting transcription must consider these components to build a coherent system.

Meetings can occur in widely varying acoustic environments (from meeting rooms to coffee shops). The realities of Internet-based communication can involve challenging network conditions, including packet loss, lag, bandwidth limitations, or audio. These can impact many different audio compression codecs used by land-line, cellular, satellite phone, and IP-based video conferencing providers. These factors significantly impact recorded audio quality and must be accounted for by meeting transcription systems designed to work in real-world environments. Traditional speech transcription solutions do not intend to deal with these issues; therefore, they must specifically build meeting transcription systems for the task. 

The primary benefit of meeting transcription is gaining actionable insights and knowledge from recorded conversations at scale. When a company is recording hundreds, thousands, or even tens or hundreds of thousands of calls, using humans to review call recordings is impractical manually. Converting these recordings into searchable text provides significant time savings and a significant expansion of capability. For instance, humans typically cannot search an audio recording, but many tools and techniques exist for searching and analyzing text.

Many companies struggle with the massive volume of meeting recordings and the sheer amount of data generated from meetings. Transcribing these recordings is an essential step in the journey, from recording a meeting to developing actionable business value. Today's AI/ML technologies have reached a quality threshold where accurately transcribing meetings, despite all their unique challenges, is finally possible. The ability for transcripts to be generated by machines without any human intervention means conversations can be examined at scale, which wasn't feasible in the era of manual call recording and call review.

One may think that meeting transcription starts with a recording. Today, systems automate to obtain quality meeting recordings without any effort or involvement by meeting participants. These systems typically connect to a company's calendar and other business systems, identifying when calls or meetings and joining them automatically. Meetings may have an "AI note taker" participant on the call, recording what was said and shown for later analysis. Some systems also integrate with a company's existing call recording or cloud recording management systems to ingest and analyze calls recorded via Sales Dialers, Video conferencing systems, or IP calling solutions, such as on-site or cloud PBXes.

Once a meeting recording is obtained, a meeting transcription solution will use a whole series of audio processing and AI/ML analysis steps to transform the raw audio into a correctly formatted and speaker-attributed conversation transcript. The following provides a brief overview of some of the steps involved in the meeting transcription process:

Media Transcoding And Stream Extraction

Meetings may be recorded in various media formats, using many different compression codecs, bit-rates, sample encodings, and other details. A system must be capable of transcoding different media formats, extracting audio and video streams, and converting these to a format suitable for downstream AI/ML analysis.

Noise Reduction And Audio Preprocessing

Conversations are sometimes in uncontrolled environments, with varying amounts of background noise, reverberation, and other audio artifacts. Before a meeting transcription system can analyze audio, it must first remove these artifacts and run a series of audio processing algorithms to improve the recorded audio quality, including pre-emphasis, dynamic range compression, and equalization.

Automatic Speech Recognition (ASR)

Meeting transcription is a complex problem for traditional ASR systems. It involves transcribing natural human conversation with unique challenges such as disfluencies, rapid-fire back-and-forth conversations, interruptions, overtalk, etc.

Speaker Diarization

Traditional ASR systems typically do not support speaker tracking (diarization); therefore, transcripts lack an attribution of statements to specific speakers. This is a significant capability in meeting transcription and involves identifying the unique characteristics in each speaker's voice, processing them into groups, and assigning statements to speakers.

Speaker Identification

Speaker identification involves going beyond determining how many speakers are involved in a conversation by attributing identities to those speakers. While traditional diarization systems may indicate that "Speaker 1 said Hello," - these systems do not know who Speaker 1 is. For this, a meeting transcription solution must implement speaker identification.

Passive Speaker Enrollment

Meetings with real-world individuals differ from remote meetings. In remote meetings, speech transcription systems can easily pick up the speaker's voice and be detected. For these systems to work in practical, real-world environments, it's essential to use a technique known as "passive speaker enrollment." This technique ensures that live meetings are transcribed and automatically attributed to participants without any human effort involved. 

Transcript Cleanup

The raw output of speech transcription systems is hard to read and requires post-processing to be usable. A transcription system should correctly be formatting numbers, dates, capitalization of names, proper nouns, and acronyms. Disfluencies and word repetitions should be removed and cleaned up for readability.

Utterance Splitting

Transcription systems must split long streams of speech into individual utterances while respecting speaker turns and ensuring the transcript attributes the speakers.

While meeting transcripts provide a searchable text version of what's in a recorded meeting, they lack fine-grained detail and context and require additional analysis and cleanup to help generate actionable insights. These transcripts go through natural language processing known as "unstructured text," meaning they lack a standardized structure or schema and are not directly machine-readable. Meeting transcripts are typically worse than unstructured written text due to the way people communicate verbally: people sometimes repeat themselves, ramble, interrupt one another, are ambiguous, are unclear, use colloquialisms, are not careful with grammar, are imprecise, use slang, are highly informal, and so on.

A company attempting to extract actionable business insights from meeting transcripts alone will likely face significant challenges due to the unstructured nature of such data. It is essential to understand that meeting transcripts are merely a first step in the journey from recording a meeting to generating actionable business value.

Conversational intelligence solutions aimed at moving past basic transcription may include the following features and functionality:

Summarization: Meeting summarization identifies discussion flow and summarizes everything into short, readable statements, often resembling meeting minutes.

Key Event Extraction: Natural Language Processing techniques extract critical events and topics of interest from a conversation.

Topic And Named Entity Extraction: Topics identified and assigned to utterances can be entities such as names, job titles, and places mentioned in the transcript that become extracted. 

Sentiment Analysis: Identifying the sentiment or attitude of a speaker in a conversation is critical when combined with named entity and topic extraction, which attributes emotions and feelings towards specific products, services, companies, etc.

Intent Extraction: Extracting information about actionable intents, such as a request for a demo or pricing, inquiry for information, proposal, etc.

These are just a few technical approaches that may be applied to meeting transcripts to generate actionable insights. Voice AI and conversational intelligence solutions are rapidly expanding in the types of analysis performed and overall sophistication of techniques when processing meeting transcriptions.

Increasingly, companies are using video calls to replace or augment traditional phone and audio-only conference calls. Video calls are often called "remote meetings" or "virtual meetings," allowing participants to join a meeting via a video call. Participants can use screen-sharing features to show products, slide decks demos, other visual content, etc.  

Meeting transcription systems should support video calls, transcribing the spoken words of the meeting participants and transcribing everything shown on screen, including remote participants sharing their screen. Meeting transcription for video calls is not a matter of duplicating what happens in audio-only meetings. Still, it requires significant additional effort to process the video part of the recording.

Today's meeting transcription solutions evolve to capture and transcribe all aspects of a meeting, including spoken words and visual information. The best meeting transcription solutions will support audio-only meetings and video calls, including additional capabilities such as visual indexing, OCR analysis of shared screens, automatic extraction of presentation slides, face detection, and more.

Meeting transcription is a rapidly evolving technology space, driven by the increasing use of call and meeting recording in business. Companies are starting to apply various natural language processing techniques to meeting transcripts, using machine learning to extract meaning, turning them into actionable insights. Video analysis and computer vision techniques will become more common, enabling meeting transcription systems to detect nonverbal communication, transcribe and index screen content, and more. Businesses will be able to apply more sophisticated techniques to meeting transcripts to turn them into actionable insights, enabling better decision-making.

Meeting transcription is becoming increasingly popular due to the growing use of call recording in business. Transcripts must go through additional processing and analysis to make them understandable and generate actionable insights. As a result, meeting transcription is becoming an essential component of broader conversational intelligence solutions to deliver business value from analyzed conversations.

As the meeting transcription market evolves, solutions will improve capturing meetings, analyzing audio and visual information, and generating actionable insights from recordings. Companies looking to embrace call and meeting transcription should look at solutions such as Hyperia, which goes far beyond basic transcription, offering capabilities such as automatic summarization, key event extraction, video call analysis, presentation slide extraction, and so on.