Skip to content

SOTA Open Source TTS

Improve loading speed & reduce memory to 1/4

Lengyuef6a35e09df6e
diff --git a/data_server/src/main.rs b/data_server/src/main.rs
index bd7c58c..b43077d 100644
--- a/data_server/src/main.rs
+++ b/data_server/src/main.rs
@@ -17,14 +17,29 @@ use text_data::{
SampleDataRequest, SampledData, Sentence, TextData,
};
+#[derive(Default, Debug, Clone)]
+pub struct RSSentence {
+ text: String,
+ phones: Vec<String>,
+ semantics: Vec<Vec<u8>>,
+}
+
+#[derive(Default, Debug, Clone)]
+pub struct RSTextData {
+ source: String,
+ name: String,
+ languages: Vec<String>,
+ sentences: Vec<RSSentence>,
+}
+
#[derive(Default)]
pub struct MyDataService {
- groups: Vec<TextData>,
+ groups: Vec<RSTextData>,
causual_sampling: bool,
weights: Vec<f32>,
}
-fn read_pb_stream<R: Read>(mut reader: BufReader<R>) -> io::Result<Vec<TextData>> {
+fn read_pb_stream<R: Read>(mut reader: BufReader<R>) -> io::Result<Vec<RSTextData>> {
let mut text_data_list = Vec::new();
let mut index = 0;
let mut total_vq_frames = 0;
@@ -47,9 +62,32 @@ fn read_pb_stream<R: Read>(mut reader: BufReader<R>) -> io::Result<Vec<TextData>
text_data.sentences.iter().for_each(|sentence| {
total_vq_frames += sentence.semantics[0].values.len();
+
+ // Check that all values are in the range 0-255
+ sentence
+ .semantics
+ .iter()
+ .for_each(|semantics| semantics.values.iter().for_each(|v| assert!(*v <= 255)));
});
- text_data_list.push(text_data);
+ text_data_list.push(RSTextData {
+ source: text_data.source.clone(),
+ name: text_data.name.clone(),
+ languages: text_data.languages.clone(),
+ sentences: text_data
+ .sentences
+ .iter()
+ .map(|sentence| RSSentence {
+ text: sentence.text.clone(),
+ phones: sentence.phones.clone(),
+ semantics: sentence
+ .semantics
+ .iter()
+ .map(|semantics| semantics.values.iter().map(|v| *v as u8).collect())
+ .collect(),
+ })
+ .collect(),
+ });
index += 1;
@@ -74,8 +112,8 @@ impl MyDataService {
// Assuming read_pb_stream is implemented and it returns an iterator over TextData
for text_data in read_pb_stream(reader)? {
- groups.push(text_data.clone());
weights.push(text_data.sentences.len() as f32); // Assuming sentences is a repeated field in TextData
+ groups.push(text_data);
}
}
@@ -87,6 +125,20 @@ impl MyDataService {
}
}
+fn cast_rs_sentence(sentence: &RSSentence) -> Sentence {
+ Sentence {
+ text: sentence.text.clone(),
+ phones: sentence.phones.clone(),
+ semantics: sentence
+ .semantics
+ .iter()
+ .map(|semantics| text_data::Semantics {
+ values: semantics.iter().map(|v| *v as u32).collect(),
+ })
+ .collect(),
+ }
+}
+
#[tonic::async_trait]
impl DataService for MyDataService {
async fn sample_data(
@@ -117,7 +169,7 @@ impl DataService for MyDataService {
return Ok(Response::new(SampledData {
name: group.name.clone(),
source: group.source.clone(),
- samples: group.sentences.clone(),
+ samples: (&group.sentences).into_iter().map(cast_rs_sentence).collect(),
}));
}
@@ -125,14 +177,14 @@ impl DataService for MyDataService {
Ok(Response::new(SampledData {
name: group.name.clone(),
source: group.source.clone(),
- samples: group.sentences[start..start + num_samples].to_vec(),
+ samples: group.sentences[start..start + num_samples].iter().map(cast_rs_sentence).collect(),
}))
} else {
let sentences_ref = group.sentences.choose_multiple(&mut rng, num_samples);
let sentences: Vec<Sentence> = sentences_ref
.into_iter()
- .cloned() // Clone each &Sentence to get Sentence
+ .map(cast_rs_sentence)
.collect();
Ok(Response::new(SampledData {