rustyverse/rustytorch
2
0
Fork 0
Code Issues Pull Requests Actions 44 Packages Projects Releases Wiki Activity

rtx-csm: emotional_speech_guide — CREMA-D vs RAVDESS firdhokk verdict

Browse Source 
8-gen bench (4 emotions × 2 corpora) at seed=42 against firdhokk
Whisper-LV3:

  target    RAVDESS              CREMA-D
  happy     happy (0.999) ✓      happy (0.999) ✓
  angry     neutral (0.92)       sad (0.99)
  fearful   happy (0.998)        fearful (0.984) ✓
  sad       angry (0.99)         fearful (0.99)

CREMA-D 2/4 vs RAVDESS 1/4. Larger / more naturalistic corpus
produces more class-pure fearful direction. Neither corpus solves
angry or sad — recipe shifts into 'vague expressivity' rather than
class-specific corners.

Practical: prefer CREMA-D when available; A/B both per emotion if
class precision matters.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
This commit is contained in:
osobh
2026-04-30 00:01:02 -07:00
co-authored by Claude Opus 4.7
parent f4d8268381
commit a5cedfb46a
69 changed files with 1026 additions and 734 deletions
Download Patch File Download Diff File Expand all files Collapse all files
crates/models/rtx-csm/src/wavlm_sv_convert.rs
+17 -10
View File
@@ -25,8 +25,8 @@
//! `candle_core::pickle::read_all` reads the `.bin` directly. No Python
//! step needed in the conversion pipeline.
use anyhow::{anyhow, Context, Result};
use candle_core::{pickle, safetensors as ct_safetensors, Tensor};
use anyhow::{Context, Result, anyhow};
use candle_core::{Tensor, pickle, safetensors as ct_safetensors};
use std::collections::HashMap;
use std::path::Path;
@@ -119,9 +119,7 @@ pub fn merge_weight_norm_auto(v: &Tensor, g: &Tensor) -> Result<Tensor> {
pub fn merge_weight_norm_dim(v: &Tensor, g: &Tensor, kept_dim: usize) -> Result<Tensor> {
let rank = v.rank();
if kept_dim >= rank {
return Err(anyhow!(
"kept_dim {kept_dim} out of range for rank {rank}"
));
return Err(anyhow!("kept_dim {kept_dim} out of range for rank {rank}"));
}
let mut norm_sq = v.sqr().context("v.sqr")?;
for axis in (0..rank).rev() {
@@ -151,8 +149,12 @@ mod tests {
let n0 = (1.0f32 + 4.0 + 9.0).sqrt();
let n1 = (16.0f32 + 25.0 + 36.0).sqrt();
let expected = [
2.0 * 1.0 / n0, 2.0 * 2.0 / n0, 2.0 * 3.0 / n0,
3.0 * 4.0 / n1, 3.0 * 5.0 / n1, 3.0 * 6.0 / n1,
2.0 * 1.0 / n0,
2.0 * 2.0 / n0,
2.0 * 3.0 / n0,
3.0 * 4.0 / n1,
3.0 * 5.0 / n1,
3.0 * 6.0 / n1,
];
for (a, b) in m.iter().zip(expected.iter()) {
assert!((a - b).abs() < 1e-6);
@@ -165,7 +167,9 @@ mod tests {
// (norm taken over axes 0,1). Per-kernel-position scale.
let device = Device::Cpu;
let v = Tensor::from_slice(
&[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0],
&[
1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0,
],
(2, 2, 3),
&device,
)
@@ -188,8 +192,11 @@ mod tests {
for j in 0..2 {
let idx = i * 6 + j * 3 + k;
let want = g_flat[k] * v_flat[idx] / norm;
assert!((m_flat[idx] - want).abs() < 1e-5,
"k={k} i={i} j={j}: got {} want {want}", m_flat[idx]);
assert!(
(m_flat[idx] - want).abs() < 1e-5,
"k={k} i={i} j={j}: got {} want {want}",
m_flat[idx]
);
}
}
}