Speech-to-Text

Profanity Filtering Expanded Language Support

Profanity Filtering Gets Expanded Language Support

Profanity filtering now supports 6 additional languages beyond English, giving you content moderation capabilities across your global user base. Available on monolingual models for:

Newly Supported:

• German (de)
• Swiss German (de-CH)
• Polish (pl)
• Portuguese (pt, pt-BR, pt-PT)
• Spanish (es, es-419)
• Swedish (sv, sv-SE)

Existing English Support:

• en, en-US, en-AU, en-GB, en-NZ, en-IN

This expansion lets you deploy consistent content policies across international markets without building custom filtering logic.

Smart Formatting Improvements

We’ve resolved several high-impact formatting edge cases that were causing transcription accuracy issues in production environments:

Improved Entity Formatting via Smart Format

Email Transcription Improvements

• Fixed: 'o' characters in email addresses now transcribe correctly instead of converting to '0'
• Fixed: edge case email mentions that were being dropped entirely in specific batch processing scenarios

Certain formerly numeric-only sequences have been updated to correctly preserve all alphanumeric characters:

• Before (some entities): "my account number is a b c d zero nine" → "my account number is 09"
• After (some entities): "my account number is a b c d zero nine" → "my account number is ABCD09"

Quantity modifiers (‘single’, ‘double’, ‘triple’ + standalone character or number) are better handled via Smart Format:

• Before (some entities): "double 2" → "2"
• After (some entities): "double 2" → "22"

Special cases of ‘hundred’ or ‘a hundred’ now supported via Smart Format:

• Before (some entities): "hundred percent" → "%"
• After (some entities): "hundred percent" → "100%"

This update has gone out to all hosted streaming transcription, and will be applied to our next self-hosted release later this month.

Nova-3 Medical Streaming Significant Upgrade

We’ve just released a significant upgrade to Nova-3 Medical Streaming, bringing substantial improvements in accuracy for real-time medical transcription use cases. This update focuses specifically on our streaming model, delivering better performance across key transcription metrics.

Performance Improvements

• **11% relative reduction in Overall WER **compared to Nova-3 general streaming model
• **30% relative reduction in Overall WER **compared to Nova-2 Medical streaming model
• **2.7x improvement in Keyword Recall Rate (KRR) **compared to Nova-3 general streaming model
• Maintains industry-leading inference speed with ultra-low latency for real-time healthcare applications

Availability

The updated Nova-3 Medical streaming model is now available through our API. To access:

• Use model=nova-3-medical in your streaming API calls
• Available for hosted use
• Self-hosted deployments will be available in subsequent releases
• English only

For details on the original Nova-3 Medical release (including batch capabilities), check out the original changelog: Introducing Nova-3 Medical. For detailed information about Nova-3 Medical, please refer to our Developer Documentation.

Smart Formatting Improvements

We’ve made some improvements to Smart Formatting, enhancing both streaming finalization behavior and entity recognition performance.

Streaming Finalization Improvements

Previously, when formatting entities like dates or credit card numbers, our models would sometimes wait for additional words before finalizing the transcript—particularly if the entity seemed incomplete. For example, when someone said “nineteen seventy…” Deepgram might pause, expecting a possible follow-up like “nine” or other additional speech before finalizing the complete year.

Now, instead of potentially waiting indefinitely for additional words, our system will finalize the transcript after 3 seconds of silence, and attempt to format the entity based on the available audio. This helps ensure transcripts are returned faster and more reliably—without sacrificing too much formatting precision.

Want more control over the finalization behavior? You have two options:

• Implement logic to send a Finalize message earlier than the 3-second threshold. Reference our Finalize documentation here.
• Set no_delay=true to override formatting and force immediate finalization. NOTE: This will result in skipping formatting altogether in many cases.

Enhanced Entity Formatting

In addition to the timeout improvements, we’ve refined our Named Entity Recognition model for Smart Formatting to better identify and format:

• Date variations
• Alphanumerics (order numbers, membership IDs, prescription numbers, etc.)
• Currencies
• Payment and card information
• SSNs
• Time zones

This update is automatically applied to all streaming transcription using Smart Formatting, and is included in our Self-Hosted March 2025 Release (250331). For more details, check out our Smart Formatting documentation.

Nova-3 Multilingual General Availability - Real-Time Code-Switching

Deepgram is proud to announce the general availability of Nova-3 Multilingual, the first model of its kind able to codeswitch in real-time across 10 different languages. This revolutionary capability unlocks a host of new possibilities for global operations by processing multilingual conversations instantly with a single model—an industry-first breakthrough that changes the game for speech recognition.

Multilingual Support

• Real-time multilingual speech recognition with a truly unified speech recognition system

• Supports code-switching between 10 languages:

  • English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch

• Seamlessly handles natural language transitions without relying on explicit routing or language-specific mechanisms

• Maintains high transcription accuracy across languages while adapting to natural language transitions

• Developed through a multi-stage training process combining synthetic code-switched data at massive scale with carefully curated real-world datasets

Use Cases

Nova-3 Multilingual represents a significant breakthrough for applications in:

• Global customer support
• Emergency response (e.g., 911 calls)
• Multilingual meetings
• Retail interactions
• Healthcare settings

In high-stakes scenarios like emergency response, Nova-3 can fluidly process interactions where callers switch between languages (e.g., Spanish and English) in real time, ensuring dispatchers receive accurate, immediate transcriptions without missing critical details.

Availability

• Now available through our API
• Use model=nova-3&language=multi in your API calls
• Supports both pre-recorded and real-time streaming transcription
• Available for hosted and self-hosted use

For detailed information about Nova-3 Multilingual, please refer to our Developer Documentation.

Expanded Numerals Language Support

Deepgram is excited to announce expanded language support for numerals through our numerals and smart_format parameters, providing more comprehensive coverage for converting written numbers to numerical format across additional languages.

Expanded Language Support

• New languages added to Numerals support:
  • Danish: da, da-DK
  • Dutch: nl
  • French: fr, fr-CA
  • German: de
  • German (Switzerland): de-CH
  • Italian: it
  • Norwegian: no
  • Polish: pl
  • Portuguese: pt, pt-BR, pt-PT
  • Spanish: es, es-419
  • Swedish: sv, sv-SE

Previously supported language:

• English: en, en-US, en-AU, en-GB, en-NZ, en-IN

Feature Integration

• All newly supported languages are fully integrated with our Smart Format feature
• When using smart_format=true, numerals will automatically be applied for all supported languages
• Individual control remains available through the dedicated numerals=true parameter

Availability

The expanded Numerals support is now available through our API for use with all Deepgram speech-to-text models.

• Available for hosted and self-hosted usage.
• Compatible with both pre-recorded and real-time streaming transcription

For detailed information about our expanded Numerals or Smart Formatting support, please refer to our Developer Documentation.

Nova-3 Medical Model

Deepgram is proud to announce the release of Nova-3-Medical, our most advanced speech-to-text model designed specifically for clinical environments. Built on the foundation of Nova-3, this specialized model delivers superior transcription performance for healthcare applications.

Performance Improvements

• 63.7% reduction in word error rate (WER) compared to next-best competitor (3.44% median WER)
• 40.35% reduction in Keyword Error Rate (KER) compared to next-best competitor (6.79% KER)
• 10.6% improvement in Keyword Recall Rate (KRR) over Nova-2-Medical (93.99% KRR)
• Maintains industry-leading inference speed with ultra-low latency for real-time healthcare applications

New Features

• Self-serve customization through Keyterm Prompting

  • Instantly adapt up to 100 domain-specific terms without model retraining

• Enhanced capabilities for clinical environments

  • Improved handling of background noise typical in healthcare settings
  • Superior recognition of medication names, diagnostic terms, and procedure details
  • Accurate transcription even with far-field devices or ambient noise interference
  • Maintains exceptional accuracy in bustling clinics or hospitals with active medical equipment

Availability

Nova-3 English is now available through our API. To access:

• Use model=nova-3-medical in your API calls
• Available for hosted use
• Supports both pre-recorded and real-time streaming transcription
• Self-hosted deployments will be available in subsequent releases

For detailed information about Nova-3 Medical, please refer to our Developer Documentation.

Introducing Nova-3 - Most Advanced Speech-to-Text Model

Deepgram is proud to announce the release of Nova-3, our most advanced speech-to-text model to date. Key improvements include:

Performance Improvements

• 54.3% reduction in word error rate (WER) for streaming audio compared to competitors (6.84% median WER)
• 47.4% reduction in WER for batch processing (5.26% median WER)
• Maintains industry-leading inference speed, with latency comparable to Nova-2

New Features

• Self-serve customization through Keyterm Prompting

  • Instantly adapt up to 100 domain-specific terms without model retraining
  • Improved recognition of specialized vocabulary and technical terminology

• Enhanced capabilities for challenging audio conditions:

  • Improved handling of background noise and overlapping speech
  • Better numeric recognition
  • Real-time redaction for up to 50 entities
  • Greater word-level timestamp precision
  • Improved English formatting and paragraph structuring

Availability

Nova-3 English is now available through our API. To access:

• Use model=nova-3 in your API calls
• Available for hosted use
• Supports both pre-recorded and real-time streaming transcription
• Multilingual and self-hosted deployments will be available in subsequent releases

For detailed information about Nova-3, please refer to our Developer Documentation.

Enhanced Profanity Filtering System

What’s New

The profanity filtering system (profanity_filter=true) now includes:

• An expanded database of profane words and phrases
• Real-time filtering support for streaming audio
• Standardized asterisk replacement for filtered content

Note: Profanity Filtering is available for English-only.

Key Changes

Previously, our system would replace detected profanity with similar-sounding words. With this update, all identified profane content is now consistently replaced with asterisks, providing clearer and more predictable output.

Documentation & Support

• See our Profanity Filter documentation to learn more.
• For Self-Hosted customers, please reach out to your Deepgram contact if you’d like to use the updated feature.

Nova-2 Cantonese Language Support

model=nova-2&language=zh-HK

We’re excited to announce that we have released** Cantonese support** in our next-gen speech-to-text model, Nova-2.

See our documentation for information.

For Self-Hosted customers, please reach out to your Deepgram contact if you’d like to use the new language.

Nova-2 Air Traffic Control Model Release

model=nova-2-atc

We’re excited to announce that we have released a use case model for transcribing English **air traffic control **audio in our next-gen speech-to-text model, Nova-2.

With full pre-recorded and streaming capabilities, it’s optimized for noisy and jargon heavy air traffic control conversations.

See our documentation for information.

For Self-Hosted customers, please reach out to your Deepgram contact if you’d like to use the new model.