Add model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- sw
|
| 4 |
+
tags:
|
| 5 |
+
- whisper
|
| 6 |
+
- automatic-speech-recognition
|
| 7 |
+
- swahili
|
| 8 |
+
- nonstandard-speech
|
| 9 |
+
- kenya
|
| 10 |
+
- low-resource
|
| 11 |
+
datasets:
|
| 12 |
+
- ElizabethMwangi/swahili_afrivoice_all_domains_v0.1
|
| 13 |
+
- cdli/common_voice_swahili_small
|
| 14 |
+
- cdli/kenyan_swahili_nonstandard_speech_v1.0
|
| 15 |
+
base_model: ElizabethMwangi/whisper-large-v3-swahili-afrivoice-cv
|
| 16 |
+
license: apache-2.0
|
| 17 |
+
---
|
| 18 |
+
|
| 19 |
+
# Whisper Large V3 — Swahili NSS (AfriVoice 5 Domains + Common Voice)
|
| 20 |
+
|
| 21 |
+
Stage 2 nonstandard speech fine-tune of [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) on the CDLI Kenyan Swahili Nonstandard Speech dataset. Built on top of [ElizabethMwangi/whisper-large-v3-swahili-afrivoice-cv](https://huggingface.co/ElizabethMwangi/whisper-large-v3-swahili-afrivoice-cv), a Stage 1 checkpoint trained on AfriVoice (all 5 domains) interleaved with Common Voice Swahili.
|
| 22 |
+
|
| 23 |
+
## Training Details
|
| 24 |
+
|
| 25 |
+
- **Stage 1 base:** ElizabethMwangi/whisper-large-v3-swahili-afrivoice-cv (AfriVoice 5 domains + Common Voice)
|
| 26 |
+
- **Stage 2 dataset:** cdli/kenyan_swahili_nonstandard_speech_v1.0 (849 examples)
|
| 27 |
+
- **Language token:** sw
|
| 28 |
+
- **Full model training:** encoder, decoder, and projection updated
|
| 29 |
+
- **SpecAugment:** On
|
| 30 |
+
- **Learning rate:** 5e-5 (polynomial decay, warmup 40 steps)
|
| 31 |
+
- **Max steps:** 1000
|
| 32 |
+
- **Best checkpoint:** step 550
|
| 33 |
+
- **Output dir:** sw_nss_tune_whisper_v3_large_afrivoice_v1.2
|
| 34 |
+
|
| 35 |
+
## Results on Kenyan Swahili Nonstandard Speech (test set, best checkpoint)
|
| 36 |
+
|
| 37 |
+
**Overall, compared to unadapted baseline**
|
| 38 |
+
|
| 39 |
+
| Model | Overall WER | Avg WER |
|
| 40 |
+
|-------|-------------|---------|
|
| 41 |
+
| openai/whisper-large-v3 (no adaptation) | 0.845 | - |
|
| 42 |
+
| Stage 1 only (afrivoice-cv, no NSS tuning) | 0.498 | - |
|
| 43 |
+
| This model (adapted) | 0.328 | 0.276 |
|
| 44 |
+
|
| 45 |
+
| Metric | Value |
|
| 46 |
+
|--------|-------|
|
| 47 |
+
| Overall WER | 0.328 |
|
| 48 |
+
| Overall CER | 0.158 |
|
| 49 |
+
| Avg WER | 0.276 |
|
| 50 |
+
| Avg CER | 0.124 |
|
| 51 |
+
| Meaning preservation | 76.7% |
|
| 52 |
+
|
| 53 |
+
**By severity**
|
| 54 |
+
|
| 55 |
+
| Severity | WER | CER |
|
| 56 |
+
|----------|-----|-----|
|
| 57 |
+
| Mild | 0.21 | 0.09 |
|
| 58 |
+
| Moderate | 0.25 | 0.11 |
|
| 59 |
+
| Severe | 0.34 | 0.15 |
|