Call transcription involves converting a phone or video call into written text, in order to keep a searchable record of the conversation. According to the ElevenLabs blog, this process is most often automated using an AI speech recognition model, called speech to text, which replaces manual note-taking.
- Call transcription turns calls into searchable text, usually with automated speech-to-text.
- Businesses use it for quality control, internal search, and compliance proof under GDPR or HIPAA.
- Deferred transcription is the most accurate, while real-time transcription lets participants follow the call live.
- Consent rules vary by location, so businesses should notify all participants and apply the strictest rule.
The process works in several steps. The system first captures the audio, either by recording it for deferred transcription or by streaming it for live transcription. The model then converts the conversation into text, before formatting the result with timestamps and speaker labels, to indicate who spoke and when. Some transcriptions may also include tags for non-verbal sounds. Finally, the text is delivered to the user, who can edit and save it. Some models integrate with CRMs, these customer relationship management software, and directly file the transcriptions in the relevant customer’s record.
Why businesses use it
Organizations use transcription for several purposes. The texts are often added to an internal database, where team members can search for specific information without re-reading every exchange in full. Sales and customer service teams also use it for quality control: managers can review conversations to check that expectations are met and identify areas for improvement.
Transcription also serves as proof for companies subject to strict standards, such as HIPAA, GDPR, or SEC regulations. These documents can help demonstrate compliance with obligations during exchanges with customers and partners, particularly in the event of an audit.
Deferred or real-time transcription
Two approaches coexist. Deferred transcription, applied to a completed recording, is described by the article as the least expensive and most accurate solution, making it suitable for high-stakes meetings or compliance documentation. Real-time transcription, which relies on a streaming model, provides notes available as soon as the call ends. It can also help a non-native or hearing-impaired participant follow the discussion during the exchange.
ElevenLabs announces for its Scribe v2 Realtime model transcription in over 90 languages, with a latency of around 150 ms. The company specifies that this is only the model’s inference time, with end-to-end delay varying depending on factors such as the user’s location or the type of access point used.
Consent and legality
The legality of transcription depends on the applicable laws at the location of the call. Some countries or states only require the consent of one party, while others require that of all participants. When a call takes place between multiple jurisdictions, the article recommends applying the strictest rule. The simplest practice is to notify all participants at the start of the call, with the continuation of the conversation generally considered implied consent.
Keeping a transcription is generally legal as long as the recording was. Texts such as GDPR or CCPA may, however, limit the retention period for sensitive data. The article specifies that it does not constitute legal advice and suggests consulting a lawyer specializing in privacy to define a suitable strategy.
