Skip to content
Sign in
  • Browse all toolsFilter by category and pricing
  • Browse categoriesWriting, code, image, audio and more
  • Compare toolsPut up to three side by side
  • Read the blogGuides and comparisons

Conformer-2

Speech recognition with improved transcription of proper nouns, alphanumerics, and noisy audio

Conformer-2 is a speech recognition model from AssemblyAI that transcribes English audio with improvements in proper nouns, alphanumerics, and noise robustness. Conformer-2 serves product and development teams building applications with spoken data and was released through the AssemblyAI API; the article now points to newer model offerings.

Updated 17 Nov 20265 sections

Conformer-2 website screenshot

What is Conformer-2?

Conformer-2 is an automatic speech recognition model from AssemblyAI for converting spoken English into text. Conformer-2 is intended for product and development teams that need speech transcription as part of applications using spoken data, including generative AI workflows. The model builds on Conformer-1 and emphasizes errors that can change the meaning or usefulness of a transcript, such as incorrectly recognized names and numerical sequences.

Conformer-2 was trained on 1.1M hours of English audio and uses a model with 450M parameters. AssemblyAI reports improvements in proper noun recognition, alphanumeric transcription, and robustness to noise, rather than a significant improvement in overall word error rate. The release article also carries a notice that newer models are available and refers readers to the documentation for current offerings.

What Conformer-2 can do

Conformer-2 addresses transcription tasks where specific words or characters matter to downstream applications. The article distinguishes overall word error rate from more targeted evaluations of names, numbers, and noisy recordings. AssemblyAI reports improvements against Conformer-1 in these targeted areas and describes infrastructure changes that also reduce the time needed to process audio files.

  • Conformer-2 converts English speech into text.
  • Conformer-2 improves transcription consistency for proper nouns such as names.
  • Conformer-2 improves recognition of alphanumeric and numerical sequences.
  • Conformer-2 improves transcription robustness when audio contains noise.
  • Conformer-2 processes audio faster than Conformer-1 in the reported comparisons.
  • Conformer-2 supports an optional speech_threshold API parameter to reject files with insufficient speech.

How Conformer-2 works

Conformer-2 was released through the AssemblyAI API. The launch introduces speech_threshold, an optional API parameter that sets the minimum proportion of speech an audio file must contain to be processed. The API rejects files below the chosen threshold. The article identifies sleep podcasts, instrumental music, and empty audio files as examples where users may want to avoid unnecessary transcription.

Conformer-2 was trained using model ensembling within a noisy student-teacher training approach. Multiple teacher models generated labels for training data, exposing the student model to a wider distribution of behaviors than a single teacher could provide. AssemblyAI also increased training data and model size, while investing in serving infrastructure to improve processing speed despite the larger model.

Who Conformer-2 is for

Conformer-2 is aimed at product and development teams building applications that depend on accurate speech-to-text output. The article evaluates performance using human-labeled audio from call centers, podcasts, broadcasts, and webinars. The discussion also highlights applications where transcription errors involving names, addresses, credit card numbers, or confirmation codes can have greater consequences than ordinary word substitutions.

  • Product teams can use Conformer-2 transcripts as spoken-data input for generative AI applications.
  • Developers can use Conformer-2 for transcription workflows involving call center recordings.
  • Audio application teams can use Conformer-2 to transcribe podcasts, broadcasts, and webinars.
  • Applications handling names or numerical sequences can use Conformer-2 for the targeted recognition improvements described in the release.

Conformer-2 strengths and limitations

Conformer-2 combines targeted recognition improvements with faster processing in the comparisons reported by AssemblyAI. The article reports a 6.8% improvement in proper noun error rate and a 12.0% improvement in noise robustness at the specified test condition. The optional speech-content threshold also gives API users a way to reject unsuitable files before transcription. These results describe the release's evaluations rather than guaranteed performance for every recording.

  • Conformer-2 maintains overall word error rate parity with Conformer-1 rather than showing significant improvement on that metric.
  • Conformer-2 speed improvements vary with the duration of the audio file.
  • Conformer-2 has been followed by newer models, so the release article does not establish current model availability.

Conformer-2 FAQ

What is Conformer-2?

Conformer-2 is an automatic speech recognition model from AssemblyAI, trained on 1.1M hours of English audio. Conformer-2 builds on Conformer-1 with improvements in proper noun recognition, alphanumeric transcription, noise robustness, and processing speed while maintaining overall word error rate parity.

Who is Conformer-2 for?

Conformer-2 is for product and development teams building applications that use spoken data. The release discusses speech-to-text as a component of generative AI pipelines and evaluates audio from domains including call centers, podcasts, broadcasts, and webinars.

Does Conformer-2 have an API?

Conformer-2 was made accessible through the AssemblyAI API at launch. The release also introduces the optional speech_threshold parameter for rejecting audio with insufficient speech. The article now directs readers to documentation for current model offerings because newer models have been released.

How does Conformer-2 handle noisy audio?

Conformer-2 improves noise robustness compared with Conformer-1 in the reported evaluation using speech with added white noise. AssemblyAI reports a 12.0% improvement at a signal-to-noise ratio of 0, where the original audio and added noise have equal proportions.

What does the Conformer-2 speech_threshold parameter do?

Conformer-2 supports an optional speech_threshold API parameter that specifies the minimum proportion of speech an audio file must contain for processing. The API rejects files below that threshold, helping users control unwanted transcription costs for recordings such as instrumental music, sleep podcasts, or empty audio.

Does Conformer-2 improve overall word error rate?

Conformer-2 maintains parity with Conformer-1 in overall word error rate rather than delivering a significant improvement on that measure. The reported gains instead focus on proper noun errors, alphanumeric transcription, noise robustness, and transcription processing speed.

From the community

No reviews yet

Be the first to review this tool. What you write helps the next person choose.

More tools like Conformer-2

AssemblyAI

www.assemblyai.com

Speech-to-text, text-to-speech, and speech understanding APIs for voice applications.

Freemium

Ermine

www.ermine.ai

Local audio recording and English transcription in the browser, with audio and transcript downloads.

Freemium

Deepgram ASR

www.deepgram.com

Speech-to-text APIs, text-to-speech, and voice agents for real-time voice applications.

Freemium

VibeVoice

www.github.com

Open-source speech recognition and real-time text-to-speech models for voice AI research

Free