Scribe v2 just got an upgrade

Last updated Apr 2, 2026 • 2 minutes reading time

We are introducing built-in entity redaction, improved Indic-English transcription, a new No Verbatim mode, and expanded keyterm prompting up to 1,000 terms.

These improvements give developers more control over privacy, multilingual accuracy, transcript quality, and domain-specific precision.

New entity redaction in Scribe v2

Scribe v2 can now automatically detect and redact sensitive entities from transcripts - including names, credit card numbers, Social Security numbers, and other personally identifiable information.

Redaction happens during transcription. Sensitive data is removed before it reaches your storage or downstream systems.

Complete redaction: replaces entities with [REDACTED];
Categorized: replaces with the entity type, e.g. [CREDIT_CARD];
Enumerated: replaces with a categorized and numbered label, e.g. [CREDIT_CARD_1].

This is particularly relevant for teams in healthcare, finance, and customer support, where compliance requirements require that PII is handled before transcripts are stored or shared.

Accurate Indic-English code-switching

Across India, speakers frequently mix English with Hindi, Telugu, Kannada, and other Indic languages. Many transcription systems transliterate English words into Indic scripts, producing transcripts that do not reflect how people actually communicate.

Scribe v2 transcribes English words in English, regardless of the surrounding language. This works automatically, with no language configuration required. Whether you pass English, Hindi, or no language code at all, English words remain in Latin script.

This applies across Indic languages, not only Hindi-English. If a speaker switches between Telugu and English or Kannada and English, the English portions are transcribed correctly.

No Verbatim mode for clean transcripts

We are introducing No Verbatim mode, a transcription setting that automatically removes filler words such as "um" and "uh", along with repeated phrases and stuttering.

The result is a clean, readable transcript without manual editing or post-processing.

No Verbatim mode is well suited for meeting notes, subtitles, and any workflow where the goal is a polished written record rather than a raw capture of every sound.

Keyterm prompting expanded to 1,000

Keyterm prompting now supports up to 1,000 words and phrases per transcript, up from the previous limit of 100.

This gives teams working with large technical vocabularies, product catalogs, or domain-specific terminology more room to guide the model toward accurate transcription.

Keyterm prompting remains context-aware. The model uses surrounding audio to determine whether a keyterm applies, rather than inserting terms blindly. At 1,000 terms, this represents 10x the previous capacity.

Note: Requests with more than 100 keyterms have a minimum billable unit of 20 seconds.