What Underdog Bark 3 actually does
Underdog's Bark 3 release is an English-only automatic speech recognition family. Its base model downloads at about 163 MB; Clean and Plus are about 180 MB, while Pro is about 1.25 GB. The official model card lists CC BY 4.0 licensing and attribution details. This is a transcription release, not the similarly named speech-generation project.
Verbatim and Clean solve different problems
The base model preserves fillers in speech; Clean is intended to remove fillers and false starts for dictation. The release does not yet offer word timestamps, speaker identification or the custom word list for version 3. Those distinctions should guide the first trial rather than a general claim that one transcript looks neater.
Write two reference transcripts for one recording
Our suggested evaluation uses a short original recording with a hesitation, a corrected phrase, a name and a number. Write a verbatim reference and a separately approved clean reference before running either variant. Score base output against the verbatim version and Clean against the intended edited version. Otherwise a system can be penalized for removing words it was designed to remove, or rewarded for silently changing meaning. Preserve the recording so disputed words can be checked.
A compact download is not a memory measurement
The model card warns that packed weights expand in memory. It also notes that stock MLX audio processing can increase memory use with recording length and recommends splitting long recordings on smaller-memory Macs. Its Windows instructions use a PyTorch CUDA or CPU runtime rather than a native low-bit GPU export.
Measure useful transcription on your own files
Record first-load time, processing time and peak memory separately. Start with a short clip and then a longer recording with the same speaker and microphone. Check names, decimal values, negation and corrected statements manually. A transcript that drops one important 'not' can fail a task despite looking fluent. If you split audio, inspect the boundaries for missing or duplicated words. The vendor's reported accuracy is not an independent FyreLinkz benchmark, and we have not reproduced its timing.
Choose a variant by the final deliverable
For notes, decide how much editing is acceptable and keep a link back to the source audio. For subtitles or multi-speaker records, confirm that missing timing and speaker features can be supplied by another verified step before choosing the model. Review the repository's NOTICE and keep attribution with a distributed integration. Compare the same recording with our Phonon 2 deployment guide, rather than comparing file sizes alone. A smaller package is useful only when the complete transcription workflow meets your requirements.
Frequently asked questions
Is Bark 3 a text-to-speech model?
No. Underdog Bark 3 transcribes English audio into text; this article does not describe a speech generator.
Why can Clean output differ from the recording?
Clean is intended to remove fillers and false starts. Compare it with an approved clean reference rather than a verbatim transcript.
Does a 163 MB download mean 163 MB of runtime memory?
No. The model card describes unpacking weights and recording-length-dependent memory use. Measure the complete chosen runtime.
Sources & further reading
Primary sources checked Oct 25, 2026. Vendor statements are attributed; editorial advice is our own.
- 1
- 2
Help us keep this useful. Report a factual correction →
