rustyverse/rustytorch
2
0
Fork 0
Code Issues Pull Requests Actions 34 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/audioseal_convert.rs
+15 -14
View File
@@ -29,8 +29,8 @@
//! enough of the PyTorch pickle format to extract a `state_dict`-style nested
//! `Dict`. So we don't need a Python step in the conversion pipeline.
use anyhow::{anyhow, Context, Result};
use candle_core::{pickle, safetensors as ct_safetensors, DType, Device, Tensor};
use anyhow::{Context, Result, anyhow};
use candle_core::{DType, Device, Tensor, pickle, safetensors as ct_safetensors};
use std::collections::HashMap;
use std::path::Path;
@@ -71,9 +71,9 @@ pub fn convert_pth(
let weight_v = v_tensors
.remove(&stem)
.expect("weight_v stem present after sort");
let weight_g = g_tensors.remove(&stem).ok_or_else(|| {
anyhow!("orphan weight_v at {stem} (no matching weight_g entry)")
})?;
let weight_g = g_tensors
.remove(&stem)
.ok_or_else(|| anyhow!("orphan weight_v at {stem} (no matching weight_g entry)"))?;
let merged = merge_weight_norm(&weight_v, &weight_g)
.with_context(|| format!("merging weight_norm at {stem}"))?;
out_map.insert(format!("{stem}.weight"), merged);
@@ -82,7 +82,9 @@ pub fn convert_pth(
if !g_tensors.is_empty() {
let orphans: Vec<_> = g_tensors.keys().cloned().collect();
return Err(anyhow!("orphan weight_g entries (no matching weight_v): {orphans:?}"));
return Err(anyhow!(
"orphan weight_g entries (no matching weight_v): {orphans:?}"
));
}
let pass_count = passthrough.len();
@@ -167,12 +169,7 @@ mod tests {
fn weight_norm_merge_matches_manual() {
// Reproduce PyTorch weight_norm formula on a small (2, 3) tensor.
let device = Device::Cpu;
let v = Tensor::from_slice(
&[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0],
(2, 3),
&device,
)
.unwrap();
let v = Tensor::from_slice(&[1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0], (2, 3), &device).unwrap();
let g = Tensor::from_slice(&[2.0f32, 3.0], (2, 1), &device).unwrap();
let merged = merge_weight_norm(&v, &g).unwrap();
let merged: Vec<f32> = merged.flatten_all().unwrap().to_vec1().unwrap();
@@ -181,8 +178,12 @@ mod tests {
let n0 = (1.0f32 + 4.0 + 9.0).sqrt();
let n1 = (16.0f32 + 25.0 + 36.0).sqrt();
let expected = vec![
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 merged.iter().zip(expected.iter()) {
assert!((a - b).abs() < 1e-6, "merge mismatch: got {a}, want {b}");