README: v4-50k now default (6.85M, clarity+timing); v3 kept for rollback
Browse files
README.md
CHANGED
|
@@ -30,7 +30,7 @@ dates, temperatures, percentages, serial numbers, and a broad bank of Taiwan/wor
|
|
| 30 |
|
| 31 |
| | |
|
| 32 |
|---|---|
|
| 33 |
-
| **Parameters** |
|
| 34 |
| **Sample rate** | 24 kHz |
|
| 35 |
| **Runtime** | `onnxruntime`, **CPU‑only**, torch‑free at inference |
|
| 36 |
| **Languages** | zh‑TW (Traditional) + English + code‑mix, single voice |
|
|
@@ -49,11 +49,19 @@ dates, temperatures, percentages, serial numbers, and a broad bank of Taiwan/wor
|
|
| 49 |
| SQUIM PESQ | 3.15 | **3.31** |
|
| 50 |
| SQUIM MOS | **4.42** | 4.24 |
|
| 51 |
|
| 52 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
24 kHz clarity, a **CC0 / commercially‑clear** reference voice, and a **much larger entity‑coverage
|
| 54 |
-
corpus** (≈30k clips vs 6.6k).
|
| 55 |
-
the acoustic mel (clarity) and converges to the best held‑out intelligibility by 60k (a mid‑training 40k
|
| 56 |
-
candidate dipped to zh‑CER 0.134 before the GAN settled).
|
| 57 |
|
| 58 |
¹ `CER(generic ASR) − CER(Taiwan‑tuned Breeze‑ASR‑25)` per zh clip; `>0` ⇒ a Taiwan‑tuned recognizer
|
| 59 |
understands it better ⇒ genuine Taiwan accent present.
|
|
|
|
| 30 |
|
| 31 |
| | |
|
| 32 |
|---|---|
|
| 33 |
+
| **Parameters** | 6.85M (5.43M acoustic + 1.17M vocoder) — *v4 default; the 4.63M v3 remains in `checkpoints/`* |
|
| 34 |
| **Sample rate** | 24 kHz |
|
| 35 |
| **Runtime** | `onnxruntime`, **CPU‑only**, torch‑free at inference |
|
| 36 |
| **Languages** | zh‑TW (Traditional) + English + code‑mix, single voice |
|
|
|
|
| 49 |
| SQUIM PESQ | 3.15 | **3.31** |
|
| 50 |
| SQUIM MOS | **4.42** | 4.24 |
|
| 51 |
|
| 52 |
+
> **v4 (current default, 6.85M).** A larger acoustic model — hidden 184, 6 encoder / 7 decoder Conv‑FFN
|
| 53 |
+
> layers, ff×4, plus **contextual prosody predictors** — trained from scratch on the v3 espeak alignment
|
| 54 |
+
> with a **multi‑resolution‑STFT clarity loss** alongside the 2D mel‑GAN. Chosen as default for **clearer,
|
| 55 |
+
> more naturally‑timed** speech (preferred by ear over the 4.63M v3). Held‑out (eval_big): **pure‑zh CER
|
| 56 |
+
> 0.083** (≈ v3), overall zh‑CER 0.108 (code‑mix is the gap — v3's English‑base warm‑start still helps mixed
|
| 57 |
+
> utterances), **PESQ 3.11 / STOI 0.968 / MOS 4.41**, en‑WER 0.083. Same **CPU real‑time** envelope as v3
|
| 58 |
+
> (RTF ≈ 0.04 on 2 cores). Long text is **auto‑chunked at punctuation** (`scripts/synth_long.py`) so
|
| 59 |
+
> utterances over ~15 s don't degrade. The 4.63M v3 (zh‑CER 0.087, table below) stays in
|
| 60 |
+
> `checkpoints/inflect-micro-fastspeech-v3-30000.pt` for rollback.
|
| 61 |
+
|
| 62 |
+
The earlier 24 kHz v3 delivered a **large code‑mix gain** (0.178 → 0.092), **higher MOS** (4.24 → 4.42),
|
| 63 |
24 kHz clarity, a **CC0 / commercially‑clear** reference voice, and a **much larger entity‑coverage
|
| 64 |
+
corpus** (≈30k clips vs 6.6k). Its metrics are the right‑hand reference below.
|
|
|
|
|
|
|
| 65 |
|
| 66 |
¹ `CER(generic ASR) − CER(Taiwan‑tuned Breeze‑ASR‑25)` per zh clip; `>0` ⇒ a Taiwan‑tuned recognizer
|
| 67 |
understands it better ⇒ genuine Taiwan accent present.
|