mirror of
https://github.com/neodyland/sbv2-api.git
synced 2025-12-23 07:59:56 +00:00
Compare commits
5 Commits
commit-879
...
commit-62b
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
62ba2c802f | ||
|
|
3c8efc716c | ||
|
|
e9ced32b70 | ||
|
|
e7a1575cbc | ||
|
|
873bbb77b6 |
@@ -127,20 +127,20 @@ impl JTalkProcess {
|
||||
Ok(phone_tone_list)
|
||||
} else if tone_values.len() == 2 {
|
||||
if tone_values == hash_set![0, 1] {
|
||||
return Ok(phone_tone_list);
|
||||
Ok(phone_tone_list)
|
||||
} else if tone_values == hash_set![-1, 0] {
|
||||
return Ok(phone_tone_list
|
||||
Ok(phone_tone_list
|
||||
.iter()
|
||||
.map(|x| {
|
||||
let new_tone = if x.1 == -1 { 0 } else { 1 };
|
||||
(x.0.clone(), new_tone)
|
||||
})
|
||||
.collect());
|
||||
.collect())
|
||||
} else {
|
||||
return Err(Error::ValueError("Invalid tone values 0".to_string()));
|
||||
Err(Error::ValueError("Invalid tone values 0".to_string()))
|
||||
}
|
||||
} else {
|
||||
return Err(Error::ValueError("Invalid tone values 1".to_string()));
|
||||
Err(Error::ValueError("Invalid tone values 1".to_string()))
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -240,39 +240,43 @@ impl TTSModelHolder {
|
||||
}
|
||||
fn find_and_load_model<I: Into<TTSIdent>>(&mut self, ident: I) -> Result<bool> {
|
||||
let ident = ident.into();
|
||||
let (bytes, style_vectors) = {
|
||||
let model = self
|
||||
.models
|
||||
.iter()
|
||||
.find(|m| m.ident == ident)
|
||||
.ok_or(Error::ModelNotFoundError(ident.to_string()))?;
|
||||
if model.vits2.is_some() {
|
||||
return Ok(true);
|
||||
}
|
||||
(model.bytes.clone().unwrap(), model.style_vectors.clone())
|
||||
};
|
||||
self.unload(ident.clone());
|
||||
let s = model::load_model(&bytes, false)?;
|
||||
if let Some(max) = self.max_loaded_models {
|
||||
if self.models.iter().filter(|x| x.vits2.is_some()).count() >= max {
|
||||
self.unload(self.models.first().unwrap().ident.clone());
|
||||
}
|
||||
}
|
||||
self.models.push(TTSModel {
|
||||
bytes: Some(bytes.to_vec()),
|
||||
vits2: Some(s),
|
||||
style_vectors,
|
||||
ident: ident.clone(),
|
||||
});
|
||||
let model = self
|
||||
// Locate target model entry
|
||||
let target_index = self
|
||||
.models
|
||||
.iter()
|
||||
.find(|m| m.ident == ident)
|
||||
.position(|m| m.ident == ident)
|
||||
.ok_or(Error::ModelNotFoundError(ident.to_string()))?;
|
||||
if model.vits2.is_some() {
|
||||
|
||||
// Already loaded
|
||||
if self.models[target_index].vits2.is_some() {
|
||||
return Ok(true);
|
||||
}
|
||||
Err(Error::ModelNotFoundError(ident.to_string()))
|
||||
|
||||
// Get bytes to build a Session
|
||||
let bytes = self.models[target_index]
|
||||
.bytes
|
||||
.clone()
|
||||
.ok_or(Error::ModelNotFoundError(ident.to_string()))?;
|
||||
|
||||
// Enforce max loaded models by evicting a different loaded model's session, not removing the entry
|
||||
if let Some(max) = self.max_loaded_models {
|
||||
let loaded_count = self.models.iter().filter(|m| m.vits2.is_some()).count();
|
||||
if loaded_count >= max {
|
||||
if let Some(evict_index) = self
|
||||
.models
|
||||
.iter()
|
||||
.position(|m| m.vits2.is_some() && m.ident != ident)
|
||||
{
|
||||
// Drop only the session to free memory; keep bytes/style for future reload
|
||||
self.models[evict_index].vits2 = None;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Build and set session in-place for the target model
|
||||
let s = model::load_model(&bytes, false)?;
|
||||
self.models[target_index].vits2 = Some(s);
|
||||
Ok(true)
|
||||
}
|
||||
|
||||
/// Get style vector by style id and weight
|
||||
|
||||
@@ -173,8 +173,15 @@ pub fn parse_text_blocking(
|
||||
}
|
||||
|
||||
pub fn array_to_vec(audio_array: Array3<f32>) -> Result<Vec<u8>> {
|
||||
// If SBV2_FORCE_STEREO is set ("1"/"true"), duplicate mono to stereo
|
||||
let force_stereo = std::env::var("SBV2_FORCE_STEREO")
|
||||
.ok()
|
||||
.map(|v| matches!(v.as_str(), "1" | "true" | "TRUE" | "True"))
|
||||
.unwrap_or(false);
|
||||
|
||||
let channels: u16 = if force_stereo { 2 } else { 1 };
|
||||
let spec = WavSpec {
|
||||
channels: 1,
|
||||
channels,
|
||||
sample_rate: 44100,
|
||||
bits_per_sample: 32,
|
||||
sample_format: SampleFormat::Float,
|
||||
@@ -183,8 +190,16 @@ pub fn array_to_vec(audio_array: Array3<f32>) -> Result<Vec<u8>> {
|
||||
let mut writer = WavWriter::new(&mut cursor, spec)?;
|
||||
for i in 0..audio_array.shape()[0] {
|
||||
let output = audio_array.slice(s![i, 0, ..]).to_vec();
|
||||
for sample in output {
|
||||
writer.write_sample(sample)?;
|
||||
if force_stereo {
|
||||
for sample in output {
|
||||
// Write to Left and Right channels
|
||||
writer.write_sample(sample)?;
|
||||
writer.write_sample(sample)?;
|
||||
}
|
||||
} else {
|
||||
for sample in output {
|
||||
writer.write_sample(sample)?;
|
||||
}
|
||||
}
|
||||
}
|
||||
writer.finalize()?;
|
||||
|
||||
Reference in New Issue
Block a user