fine tuning for first and last names?

[silvacarl2]

we want to fine tune on the 12M first and last names we have to get the accuracy up as high as possible.

we would be willing to share this with anyone that might want to help us.

question: is it best to keep the first and last names together as one audio file for fine tuning? or is it better to break them up into first names audio files and last names audio files for fine tuning?

carl

[phineas-pta]

why don't you use prompt to add new vocabulary ?

[silvacarl2]

we use initial prompt now. the issue is if they are calling in for the first time, there is no way to know what their name could be. so it has to just take the recording and do the best it can with it.

which is about 82% accurate.

since we have 12M samples with ground truth, we thoguht we could get that number much higher.

thoughts?

[dgoryeo]

@silvacarl2 have you considered WhisperBiasing? I'm not sure if that requires the 12M names to be labeled with corresonding sound in a dataset.

[silvacarl2]

NICE!!!!!!!!!!!!!!!!!

p.s. if you would like to assist us on this we can pay for assistance. 8-)

[dgoryeo]

Haha -- @silvacarl2 I was actually looking for a small gig, my planned gig was just delayed by 1 month.

[silvacarl2]

[email protected]
or
csilva@sphinxmedtech,com

[silvacarl2]

Ok, email me if you have time/are interested. We have 14M names to test with. From: dgoryeo ***@***.***> Sent: Monday, September 11, 2023 9:05 AM To: openai/whisper ***@***.***> Cc: Carl Silva ***@***.***>; Mention ***@***.***> Subject: Re: [openai/whisper] fine tuning for first and last names? (Discussion #1338) Haha -- @silvacarl2 <https://t.sidekickopen21.com/Ctc/ZT+23284/cN-F804/Jks2-6qcW69sMD-6lZ3m_W5sm BLc1RcWvFW6cMDM71_TQY6W98nC5M43CqrfN39fksH1nFrMW64345c3gmchZN2qxBDL_qWK_MQdR Bj7CMp2V-kD6X2V5MHBN2DnbmpCx7R8W6WcwVc8KLtk8W3kr-Zb8FQZ8bVlL6p022TpxYW2J2sN1 29_bwQW1pd1T05tBxgBW3FfDJx78nkLSVN_btf670BWDW3CkJG682qgPJW6KNZRG1KpM8QVMfLs_ 7lPLT2Vt0-Dg5T7Rssf8B4y1K04> I was actually looking for a small gig, my planned gig was just delayed by 1 month. - Reply to this email directly, view it on GitHub <https://t.sidekickopen21.com/Ctc/ZT+23284/cN-F804/Jll2-6qcW7Y8-PT6lZ3p6W7wy 2cW2VpQmdW5GSNq61ynvqVW81c5708lfRwXW8knxQ_2MQf5PW5QX1X04-NdvSW64GWrj92L4tFW3 RL-7F6cyLQvW9chbvW1qFVVHW8CdFZ14TqZcYN8Nd0psJ1s10W90ZHs_96KtbMW8TkL3F7xsZnhW 660xP37PDpBKW3Q_vxq5-xWqXVbLSss6ktMPVW49XT5b8GWPGVW5BnDBh8dskT5W6QyVtP5QlspG W1SfThG86B2PZW7xT_Zf2fdBl5W2C0Z4-8xzCyNW2JX5XX4-XGGHW3-RLBL4H4L57W6tDJVZ2Zv5 R1W2BYgW-496jBvW18t1XZ4RmqCxf2r47T004> , or unsubscribe <https://t.sidekickopen21.com/Ctc/ZT+23284/cN-F804/JlF2-6qcW8wLKSR6lZ3mTW8Mq tDj2GXKt-W4Nfz-b4F8Z8-W1-qlxR8-927tW56Wlp843XvhmW7x0S0g8VsPd8W4V4Cd01flnLJW5 z0DWC1dLJLxW37vJDr7NptmjW81Xhdk6L3MfmW2n_wBc577tJlV9dTgt7f0lLTW49j8-b7QlWqTW 9cm3z723gyM0W43crzP7m__ZnN6jL2WpsKy1RW5PQMXr8BqpLMW3VqYk38THVZyW4lt-C41LX2_B W6tH3222Z5TtLW683n5-81QyCBW2P8H-04JT6YYW16S0KY1SrFr8N7czVQ12TCNLW9hSXP25fMdC pW2-xNT-6Tqm6rW5y30lS2hc-wnW5BxJxK13y55jW1ctbZ85NmHb4f7R3q3z04> . You are receiving this because you were mentioned. <https://github.com/notifications/beacon/ABAGP423JIHKPERBWVMVIBTXZ4ZBJA5CNFS M6AAAAAAX6RVELCWGG33NNVSW45C7OR4XAZNRIRUXGY3VONZWS33OINXW23LFNZ2KUY3PNVWWK3T UL5UWJTQANJMOA.gif> Message ID: ***@***.*** ***@***.***> > <https://t.sidekickopen21.com/Cto/ZT+23284/cN-F804/R5Q8b45TGN5n3ZXc2fDy8W3C6 vpw1ZkpK2W21hk_G1X07DjW1GysB-3GGxknW3LFK3M1N4hVGW1Y-bx91Q44K-f1Z0grK12>