Skip to content
Sign in
Artwork for My Weird Prompts
My Weird Prompts · Wednesday · 23 min

Why Extra Voice Hours Settle Bilingual Speech

Fine-tuning a massive speech model like Parakeet TDT on personal speech can yield dramatic gains with modest datasets, dropping code-switched word error rates from 13.7% to under 5%. How data maturity stages work, why repeated vocabulary adds acoustic resilience rather than redundant weight, and how models handle Hebrew-English code-switching without regressing on base English recognition are explored through scaling curves, encoder-freezing experiments, and optimization mechanics. Key sources: • NVIDIA Parakeet TDT 0.6B v2 model card, created 2025-04-15, last modified 2026-06-29 — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2 • Adapting OpenAI's Whisper for Speech Recognition on Code-Switch Mandarin-English SEAME and ASRU2019 Datasets (full text read, pp. 3 — https://arxiv.org/abs/2311.17382 • How much speech data is necessary for ASR in African languages? An evaluation of data scaling in Kinyarwanda and Kikuyu (full text read, pp. 3 — https://arxiv.org/abs/2510.07221v1 Episode 6332 — myweirdprompts.com/n/6332 (permanent short link: myweirdprompts.com/765258)

0:00 · Intro-23:10
Play in
Transcript file

transcript

No transcript — this publisher did not publish one.

show notes

Fine-tuning a massive speech model like Parakeet TDT on personal speech can yield dramatic gains with modest datasets, dropping code-switched word error rates from 13.7% to under 5%. How data maturity stages work, why repeated vocabulary adds acoustic resilience rather than redundant weight, and how models handle Hebrew-English code-switching without regressing on base English recognition are explored through scaling curves, encoder-freezing experiments, and optimization mechanics.

Key sources:
• NVIDIA Parakeet TDT 0.6B v2 model card, created 2025-04-15, last modified 2026-06-29 — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2
• Adapting OpenAI's Whisper for Speech Recognition on Code-Switch Mandarin-English SEAME and ASRU2019 Datasets (full text read, pp. 3 — https://arxiv.org/abs/2311.17382
• How much speech data is necessary for ASR in African languages? An evaluation of data scaling in Kinyarwanda and Kikuyu (full text read, pp. 3 — https://arxiv.org/abs/2510.07221v1

Episode 6332 — myweirdprompts.com/n/6332 (permanent short link: myweirdprompts.com/765258)

chapters

7 chapters

more episodes

All episodes