Which voice model should you use?
To talk and get what you meant Bark 3.0 Clean 6.1in 100 wrong off the cuff, against what people meant
Every word, word for word Fluid Speech Mini 6.6in 100 wrong averaged over seven public sets
In another language Parakeet v3 8.1in 100 wrong averaged, and 25 languages
Leaderboard transcripts keep every “um”, so Bark Clean, which leaves them out on purpose, scores worse on those sets by design.
Words wrong, word for word
Fluid Speech Mini gets the fewest words wrong: 6.6 in 100, averaged over seven public sets. The old engine got 23.3.
Every kind of speech
| Speech | Bark 3.0 Clean | Fluid Speech Mini | Parakeet v3 | Parakeet, Heyra 0.3.10 |
|---|---|---|---|---|
| Off the cuff, meantDisfluencySpeech | 6.1 | 17.3 | 17.5 | 17.8 |
| MeetingsAMI | 17.8 | 12.1 | 11.9 | 34.3 |
| Earnings callsEarnings-22 | 19.5 | 10.3 | 16.6 | 33.3 |
| PodcastsGigaSpeech | 12.8 | 10.2 | 10.7 | 10.9 |
| Finance callsSPGISpeech | 5.6 | 5.1 | 6.7 | 66.5 |
| ParliamentVoxPopuli | 4.2 | 3.1 | 4.8 | 7.2 |
| AudiobooksLibriSpeech clean | 3.8 | 2.5 | 2.3 | 3.2 |
| Audiobooks, hardLibriSpeech other | 5.6 | 2.9 | 4.0 | 7.4 |
No one model wins everywhere: the best in each kind of speech is in gold.
Side by side
Off the cuff
SaidWell, what about, uh, papyrus, you know, made out of bam-, you know, bamboo stuff, from the from the banks of the Nile.
Parakeet v3Well, what about uh peppers, you know, made out of bam, you know, bamboo stuff from the from the banks of the now?
Bark 3.0 Cleanwhat about, pepperice, made out of bamboo stuff from the banks of the now?
Meeting
SaidIt should be a battery because uh uh remote controller should be like you you you should take it t to wherever you want…
Parakeet v3it should be a battery because uh remote controller should be like you you you should take it to to wherever you want…
Bark 3.0 CleanIt should be a battery because the remote control should be like you should take it to wherever you want…
Finance call
Saidthe capacity of the existing tank houses that we have with our improvement program, so that's an important part of this.
0.3.10Nothing
0.3.11the the capacity of the existing tank houses that we have uh with our improvement program. So that's an important part of this.
Under the hood
Engine
0.3.10
23.3
0.3.11
8.1
0.3.10 returned nothing for 60 of 100 finance-call clips.
Off the cuff, meant
Parakeet v3
17.5
Bark Clean
6.1
Against transcripts without “um”s and repeats.
7-minute take
60 s pieces
30.8
15 s pieces
10.9
Bark Clean
5.3
0.3.11 cuts long takes into 15 s pieces.
Text arrives about half a second after you let go: median 0.5 s for a 2–10 s take, 95 % under 1 s.
Method
- 100 clips each: LibriSpeech, AMI, Earnings-22, GigaSpeech, SPGISpeech, VoxPopuli, DisfluencySpeech.
- sherpa-onnx 1.13.8, int8, as in Heyra. Lower-cased, no punctuation.