Dense Slides: Optimizing inference for voice models in production - Philip Kiely, Baseten

Source Video

Optimizing inference for voice models in production - Philip Kiely, Baseten

Method

This deck is slide-only. The existing captured video frame set supplies candidate frames, then local OpenCV rejects sponsor/title/speaker-only frames, crops visible slide surfaces, deduplicates, and saves the cropped slide images.

Cropped Visible Slides

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: agent_vision.

Slide text:

Agenda

1. TTS model architecture

2. TTS performance metrics

3. Orpheus TTS optimization techniques

4. Orpheus TTS performance benchmarks

5. Infrastructure and client code

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.96
Text source: advanced OCR rapidocr-live/full.
OCR decision: ready — small code panel and compact technical text likely better handled by OCR

Slide text:

AIE VCanopyLabs Example: OrpheusTTS ·Llama3.23Bbackbone.Extendedcontextlength Increasedvocabsizefor speech-specifictokens withRoPEscaling "bos_token_td':128908, "head_din':128, ax_positlon_enbedd ngs:131072, "nun_htdden_layers':28, "num_key_value_heads:8, "attentton_dropout": "eos_token_（d:12n01, "htdden_act":"stlu', "htdden_s（ze°:372, "initializer_range:e.92 "internedtate_slze:8192, "nua_attention_heads:24, "pretratntng_tp:1, ee1.:ad11apoo "rope_scaling":( "attention_blas:false, res_norn_eps:le-85 "factor:32. "htgh_freq_factor':4.0, "low_freq_factor':1.9, e11.,adA1do "ortginal_max_positlon_enbeddings:8192

"tle_vord_enbeddings':trd "rope_theta':5

"transformers_verston':"4.47.0*

vocab_stze:156939

TRL Microsoft smol?

Hidden Non-Slide Evidence

`slide-001.jpg` — title_card confidence 0.97; speaker intro card

Classification audit: raw/sources/slide-ai-classification/dense/gmTHs5T_YAE/audit.json