Parakeet Translates French Audio (thoth-app.com)

Parakeet has launched an advanced transcription model named TDT v3, which boasts remarkable speed and accuracy for English audio, nearly matching the performance of Whisper Large V3 Turbo. However, when tested on French recordings, it surprisingly outputs clean English translations instead of transcriptions, a behavior attributed to its lack of explicit language conditioning. Unlike Whisper, which utilizes a language token to guide its transcription process, Parakeet learns from acoustic patterns without language identity signals, leading it to involuntarily translate when faced with complex or spontaneous speech. This phenomenon poses significant implications for the AI/ML community, particularly in the fields of transcription and language processing. Parakeet’s performance highlights the challenges faced by AI models in accurately interpreting and outputting non-English languages, with results showing that it produced English translations for up to 31.3% of a private French-language interview. Developers are advised to use Whisper models for non-English languages, as they maintain accuracy without the mixed-language output observed with Parakeet. This distinction reinforces the need for precise language conditioning in AI transcription to meet the demands of diverse linguistic contexts effectively.