feat: Add Kokoros backend (#9212)

Signed-off-by: Richard Palethorpe <io@richiejp.com>

.github/workflows/test-extra.yml

@@ -31,6 +31,7 @@ jobs:
       llama-cpp-quantization: ${{ steps.detect.outputs.llama-cpp-quantization }}
       acestep-cpp: ${{ steps.detect.outputs.acestep-cpp }}
       voxtral: ${{ steps.detect.outputs.voxtral }}
+      kokoros: ${{ steps.detect.outputs.kokoros }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@v6
@@ -528,3 +529,25 @@ jobs:
       - name: Test voxtral
         run: |
           make --jobs=5 --output-sync=target -C backend/go/voxtral test
+  tests-kokoros:
+    needs: detect-changes
+    if: needs.detect-changes.outputs.kokoros == 'true' || needs.detect-changes.outputs.run-all == 'true'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Clone
+        uses: actions/checkout@v6
+        with:
+          submodules: true
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y build-essential cmake pkg-config protobuf-compiler clang libclang-dev
+          sudo apt-get install -y espeak-ng libespeak-ng-dev libsonic-dev libpcaudio-dev libopus-dev libssl-dev
+          curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
+          echo "$HOME/.cargo/bin" >> $GITHUB_PATH
+      - name: Build kokoros
+        run: |
+          make -C backend/rust/kokoros kokoros-grpc
+      - name: Test kokoros
+        run: |
+          make -C backend/rust/kokoros test

.gitmodules

@@ -1,3 +1,6 @@
 [submodule "docs/themes/hugo-theme-relearn"]
 	path = docs/themes/hugo-theme-relearn
 	url = https://github.com/McShelby/hugo-theme-relearn.git
+[submodule "backend/rust/kokoros/sources/Kokoros"]
+	path = backend/rust/kokoros/sources/Kokoros
+	url = https://github.com/lucasjinreal/Kokoros

Makefile

@@ -1,5 +1,5 @@
 # Disable parallel execution for backend builds
-.NOTPARALLEL: backends/diffusers backends/llama-cpp backends/outetts backends/piper backends/stablediffusion-ggml backends/whisper backends/faster-whisper backends/silero-vad backends/local-store backends/huggingface backends/rfdetr backends/kitten-tts backends/kokoro backends/chatterbox backends/llama-cpp-darwin backends/neutts build-darwin-python-backend build-darwin-go-backend backends/mlx backends/diffuser-darwin backends/mlx-vlm backends/mlx-audio backends/mlx-distributed backends/stablediffusion-ggml-darwin backends/vllm backends/vllm-omni backends/moonshine backends/pocket-tts backends/qwen-tts backends/faster-qwen3-tts backends/qwen-asr backends/nemo backends/voxcpm backends/whisperx backends/ace-step backends/acestep-cpp backends/fish-speech backends/voxtral backends/opus backends/trl backends/llama-cpp-quantization
+.NOTPARALLEL: backends/diffusers backends/llama-cpp backends/outetts backends/piper backends/stablediffusion-ggml backends/whisper backends/faster-whisper backends/silero-vad backends/local-store backends/huggingface backends/rfdetr backends/kitten-tts backends/kokoro backends/chatterbox backends/llama-cpp-darwin backends/neutts build-darwin-python-backend build-darwin-go-backend backends/mlx backends/diffuser-darwin backends/mlx-vlm backends/mlx-audio backends/mlx-distributed backends/stablediffusion-ggml-darwin backends/vllm backends/vllm-omni backends/moonshine backends/pocket-tts backends/qwen-tts backends/faster-qwen3-tts backends/qwen-asr backends/nemo backends/voxcpm backends/whisperx backends/ace-step backends/acestep-cpp backends/fish-speech backends/voxtral backends/opus backends/trl backends/llama-cpp-quantization backends/kokoros
 
 GOCMD=go
 GOTEST=$(GOCMD) test
@@ -431,6 +431,7 @@ prepare-test-extra: protogen-python
 	$(MAKE) -C backend/python/whisperx
 	$(MAKE) -C backend/python/ace-step
 	$(MAKE) -C backend/python/trl
+	$(MAKE) -C backend/rust/kokoros kokoros-grpc
 
 test-extra: prepare-test-extra
 	$(MAKE) -C backend/python/transformers test
@@ -451,6 +452,7 @@ test-extra: prepare-test-extra
 	$(MAKE) -C backend/python/whisperx test
 	$(MAKE) -C backend/python/ace-step test
 	$(MAKE) -C backend/python/trl test
+	$(MAKE) -C backend/rust/kokoros test
 
 DOCKER_IMAGE?=local-ai
 IMAGE_TYPE?=core
@@ -586,6 +588,9 @@ BACKEND_MLX_DISTRIBUTED = mlx-distributed|python|./|false|true
 BACKEND_TRL = trl|python|.|false|true
 BACKEND_LLAMA_CPP_QUANTIZATION = llama-cpp-quantization|python|.|false|true
 
+# Rust backends
+BACKEND_KOKOROS = kokoros|rust|.|false|true
+
 # Helper function to build docker image for a backend
 # Usage: $(call docker-build-backend,BACKEND_NAME,DOCKERFILE_TYPE,BUILD_CONTEXT,PROGRESS_FLAG,NEEDS_BACKEND_ARG)
 define docker-build-backend
@@ -644,12 +649,13 @@ $(eval $(call generate-docker-build-target,$(BACKEND_ACESTEP_CPP)))
 $(eval $(call generate-docker-build-target,$(BACKEND_MLX_DISTRIBUTED)))
 $(eval $(call generate-docker-build-target,$(BACKEND_TRL)))
 $(eval $(call generate-docker-build-target,$(BACKEND_LLAMA_CPP_QUANTIZATION)))
+$(eval $(call generate-docker-build-target,$(BACKEND_KOKOROS)))
 
 # Pattern rule for docker-save targets
 docker-save-%: backend-images
 	docker save local-ai-backend:$* -o backend-images/$*.tar
 
-docker-build-backends: docker-build-llama-cpp docker-build-rerankers docker-build-vllm docker-build-vllm-omni docker-build-transformers docker-build-outetts docker-build-diffusers docker-build-kokoro docker-build-faster-whisper docker-build-coqui docker-build-chatterbox docker-build-vibevoice docker-build-moonshine docker-build-pocket-tts docker-build-qwen-tts docker-build-fish-speech docker-build-faster-qwen3-tts docker-build-qwen-asr docker-build-nemo docker-build-voxcpm docker-build-whisperx docker-build-ace-step docker-build-acestep-cpp docker-build-voxtral docker-build-mlx-distributed docker-build-trl docker-build-llama-cpp-quantization
+docker-build-backends: docker-build-llama-cpp docker-build-rerankers docker-build-vllm docker-build-vllm-omni docker-build-transformers docker-build-outetts docker-build-diffusers docker-build-kokoro docker-build-faster-whisper docker-build-coqui docker-build-chatterbox docker-build-vibevoice docker-build-moonshine docker-build-pocket-tts docker-build-qwen-tts docker-build-fish-speech docker-build-faster-qwen3-tts docker-build-qwen-asr docker-build-nemo docker-build-voxcpm docker-build-whisperx docker-build-ace-step docker-build-acestep-cpp docker-build-voxtral docker-build-mlx-distributed docker-build-trl docker-build-llama-cpp-quantization docker-build-kokoros
 
 ########################################################
 ### Mock Backend for E2E Tests

backend/Dockerfile.rust

@@ -0,0 +1,39 @@
+ARG BASE_IMAGE=ubuntu:24.04
+
+FROM ${BASE_IMAGE} AS builder
+ARG BACKEND=kokoros
+ENV DEBIAN_FRONTEND=noninteractive
+ARG TARGETARCH
+ARG TARGETVARIANT
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        build-essential \
+        git ccache \
+        ca-certificates \
+        make cmake wget \
+        curl unzip \
+        clang \
+        pkg-config \
+        libssl-dev \
+        espeak-ng libespeak-ng-dev \
+        libsonic-dev libpcaudio-dev \
+        libopus-dev \
+        protobuf-compiler && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Install Rust
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
+ENV PATH="/root/.cargo/bin:${PATH}"
+
+COPY . /LocalAI
+
+RUN git config --global --add safe.directory /LocalAI
+
+RUN make -C /LocalAI/backend/rust/${BACKEND} build
+
+FROM scratch
+ARG BACKEND=kokoros
+
+COPY --from=builder /LocalAI/backend/rust/${BACKEND}/package/. ./

backend/index.yaml

@@ -469,6 +469,26 @@
     nvidia-cuda-13: "cuda13-kokoro"
     nvidia-cuda-12: "cuda12-kokoro"
     nvidia-l4t-cuda-12: "nvidia-l4t-arm64-kokoro"
+- &kokoros
+  icon: https://avatars.githubusercontent.com/u/166769057?v=4
+  description: |
+    Kokoros is a pure Rust TTS backend using the Kokoro ONNX model (82M parameters).
+    It provides fast, high-quality text-to-speech with streaming support, built on
+    ONNX Runtime for efficient CPU inference. Supports English, Japanese, Mandarin
+    Chinese, and German.
+  urls:
+    - https://huggingface.co/hexgrad/Kokoro-82M
+    - https://github.com/lucasjinreal/Kokoros
+  tags:
+    - text-to-speech
+    - TTS
+    - Rust
+    - ONNX
+  license: apache-2.0
+  alias: "kokoros"
+  name: "kokoros"
+  capabilities:
+    default: "cpu-kokoros"
 - &coqui
   urls:
     - https://github.com/idiap/coqui-ai-TTS
@@ -2043,6 +2063,21 @@
   uri: "quay.io/go-skynet/local-ai-backends:master-metal-darwin-arm64-kokoro"
   mirrors:
     - localai/localai-backends:master-metal-darwin-arm64-kokoro
+## kokoros (Rust)
+- !!merge <<: *kokoros
+  name: "kokoros-development"
+  capabilities:
+    default: "cpu-kokoros-development"
+- !!merge <<: *kokoros
+  name: "cpu-kokoros"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-cpu-kokoros"
+  mirrors:
+    - localai/localai-backends:latest-cpu-kokoros
+- !!merge <<: *kokoros
+  name: "cpu-kokoros-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-cpu-kokoros"
+  mirrors:
+    - localai/localai-backends:master-cpu-kokoros
 ## faster-whisper
 - !!merge <<: *faster-whisper
   name: "faster-whisper-development"

backend/rust/kokoros/.gitignore

@@ -0,0 +1,3 @@
+/target/
+/proto/
+/package/

backend/rust/kokoros/Cargo.toml

@@ -0,0 +1,26 @@
+[package]
+name = "kokoros-grpc"
+version = "0.1.0"
+edition = "2021"
+
+[[bin]]
+name = "kokoros-grpc"
+path = "src/main.rs"
+
+[dependencies]
+kokoros = { path = "sources/Kokoros/kokoros" }
+
+tonic = "0.13"
+prost = "0.13"
+tokio = { version = "1", features = ["full"] }
+tokio-stream = "0.1"
+clap = { version = "4", features = ["derive"] }
+tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["env-filter"] }
+
+[build-dependencies]
+tonic-build = "0.13"
+
+[features]
+default = ["cpu"]
+cpu = ["kokoros/cpu"]

backend/rust/kokoros/Makefile

@@ -0,0 +1,25 @@
+CURRENT_DIR=$(abspath ./)
+
+.PHONY: kokoros-grpc
+kokoros-grpc:
+	mkdir -p $(CURRENT_DIR)/proto
+	cp $(CURRENT_DIR)/../../backend.proto $(CURRENT_DIR)/proto/backend.proto
+	cd $(CURRENT_DIR) && \
+		BACKEND_PROTO_PATH=$(CURRENT_DIR)/proto/backend.proto \
+		cargo build --release
+
+.PHONY: package
+package:
+	bash package.sh
+
+.PHONY: test
+test: kokoros-grpc
+	cd $(CURRENT_DIR) && cargo test
+
+.PHONY: build
+build: kokoros-grpc package
+
+.PHONY: clean
+clean:
+	cargo clean
+	rm -rf package proto

backend/rust/kokoros/build.rs

@@ -0,0 +1,15 @@
+fn main() -> Result<(), Box<dyn std::error::Error>> {
+    let proto_path = std::env::var("BACKEND_PROTO_PATH")
+        .unwrap_or_else(|_| "proto/backend.proto".to_string());
+
+    let proto_dir = std::path::Path::new(&proto_path)
+        .parent()
+        .unwrap_or(std::path::Path::new("."));
+
+    tonic_build::configure()
+        .build_server(true)
+        .build_client(false)
+        .compile_protos(&[&proto_path], &[proto_dir])?;
+
+    Ok(())
+}

backend/rust/kokoros/package.sh

@@ -0,0 +1,42 @@
+#!/bin/bash
+set -e
+
+CURDIR=$(dirname "$(realpath $0)")
+mkdir -p $CURDIR/package/lib
+
+# Copy the binary and run script
+cp -avf $CURDIR/target/release/kokoros-grpc $CURDIR/package/
+cp -rfv $CURDIR/run.sh $CURDIR/package/
+chmod +x $CURDIR/package/run.sh
+
+# Copy espeak-ng data
+if [ -d "/usr/share/espeak-ng-data" ]; then
+    cp -rf /usr/share/espeak-ng-data $CURDIR/package/
+elif [ -d "/usr/lib/x86_64-linux-gnu/espeak-ng-data" ]; then
+    cp -rf /usr/lib/x86_64-linux-gnu/espeak-ng-data $CURDIR/package/
+fi
+
+# Bundle all dynamic library dependencies
+echo "Bundling dynamic library dependencies..."
+ldd $CURDIR/target/release/kokoros-grpc | grep "=>" | awk '{print $3}' | while read lib; do
+    if [ -n "$lib" ] && [ -f "$lib" ]; then
+        cp -avfL "$lib" $CURDIR/package/lib/
+    fi
+done
+
+# Copy CA certificates for HTTPS (needed for model auto-download)
+if [ -d "/etc/ssl/certs" ]; then
+    mkdir -p $CURDIR/package/etc/ssl
+    cp -rf /etc/ssl/certs $CURDIR/package/etc/ssl/
+fi
+
+# Copy the dynamic linker
+if [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
+    cp -arfLv /lib64/ld-linux-x86-64.so.2 $CURDIR/package/lib/ld.so
+elif [ -f "/lib/ld-linux-aarch64.so.1" ]; then
+    cp -arfLv /lib/ld-linux-aarch64.so.1 $CURDIR/package/lib/ld.so
+fi
+
+echo "Packaging completed successfully"
+ls -liah $CURDIR/package/
+ls -liah $CURDIR/package/lib/

backend/rust/kokoros/run.sh

@@ -0,0 +1,23 @@
+#!/bin/bash
+set -ex
+
+CURDIR=$(dirname "$(realpath $0)")
+
+export LD_LIBRARY_PATH=$CURDIR/lib:${LD_LIBRARY_PATH:-}
+
+# SSL certificates for model auto-download
+if [ -d "$CURDIR/etc/ssl/certs" ]; then
+    export SSL_CERT_DIR=$CURDIR/etc/ssl/certs
+fi
+
+# espeak-ng data directory
+if [ -d "$CURDIR/espeak-ng-data" ]; then
+    export ESPEAK_NG_DATA=$CURDIR/espeak-ng-data
+fi
+
+# Use bundled ld.so if present (portability)
+if [ -f $CURDIR/lib/ld.so ]; then
+    exec $CURDIR/lib/ld.so $CURDIR/kokoros-grpc "$@"
+fi
+
+exec $CURDIR/kokoros-grpc "$@"

backend/rust/kokoros/sources/Kokoros

@@ -0,0 +1 @@
+Subproject commit 7089168f0ca2d8e1fcd8e523c9d75d915c6afdff

backend/rust/kokoros/src/auth.rs

@@ -0,0 +1,26 @@
+use tonic::{Request, Status};
+
+/// Returns an interceptor function if LOCALAI_GRPC_AUTH_TOKEN is set.
+pub fn make_auth_interceptor(
+) -> Option<impl Fn(Request<()>) -> Result<Request<()>, Status> + Clone> {
+    let token = std::env::var("LOCALAI_GRPC_AUTH_TOKEN").ok()?;
+    if token.is_empty() {
+        return None;
+    }
+    let expected = format!("Bearer {}", token);
+    Some(
+        move |req: Request<()>| -> Result<Request<()>, Status> {
+            let meta = req.metadata();
+            match meta.get("authorization") {
+                Some(val) => {
+                    if val.as_bytes() == expected.as_bytes() {
+                        Ok(req)
+                    } else {
+                        Err(Status::unauthenticated("invalid token"))
+                    }
+                }
+                None => Err(Status::unauthenticated("missing authorization")),
+            }
+        },
+    )
+}

backend/rust/kokoros/src/main.rs

@@ -0,0 +1,53 @@
+use clap::Parser;
+use tonic::transport::Server;
+
+mod auth;
+mod service;
+
+pub mod backend {
+    tonic::include_proto!("backend");
+}
+
+#[derive(Parser, Debug)]
+#[command(name = "kokoros-grpc")]
+struct Cli {
+    /// gRPC listen address (host:port)
+    #[arg(long, default_value = "localhost:50051")]
+    addr: String,
+}
+
+#[tokio::main]
+async fn main() -> Result<(), Box<dyn std::error::Error>> {
+    tracing_subscriber::fmt()
+        .with_writer(std::io::stderr)
+        .with_ansi(false)
+        .without_time()
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("info")),
+        )
+        .init();
+
+    let cli = Cli::parse();
+    let addr = cli.addr.parse()?;
+
+    tracing::info!("Starting kokoros gRPC server on {}", addr);
+
+    let mut builder = Server::builder();
+
+    if let Some(interceptor) = auth::make_auth_interceptor() {
+        tracing::info!("Bearer token authentication enabled");
+        let svc = backend::backend_server::BackendServer::with_interceptor(
+            service::KokorosService::default(),
+            interceptor,
+        );
+        builder.add_service(svc).serve(addr).await?;
+    } else {
+        let svc = backend::backend_server::BackendServer::new(service::KokorosService::default())
+            .max_decoding_message_size(50 * 1024 * 1024)
+            .max_encoding_message_size(50 * 1024 * 1024);
+        builder.add_service(svc).serve(addr).await?;
+    }
+
+    Ok(())
+}

backend/rust/kokoros/src/service.rs

@@ -0,0 +1,652 @@
+use std::sync::{Arc, Mutex};
+use tokio::sync::Mutex as TokioMutex;
+use tokio_stream::wrappers::ReceiverStream;
+use tonic::{Request, Response, Status};
+
+use kokoros::tts::koko::TTSKoko;
+
+use crate::backend;
+use crate::backend::backend_server::Backend;
+
+/// Write f32 samples as a standard 44-byte PCM 16-bit WAV file.
+/// LocalAI's audio pipeline assumes this exact header layout.
+fn write_pcm16_wav(
+    path: &str,
+    samples: &[f32],
+    sample_rate: u32,
+) -> Result<(), Box<dyn std::error::Error>> {
+    use std::fs::File;
+    use std::io::Write;
+
+    let num_samples = samples.len() as u32;
+    let data_size = num_samples * 2; // 16-bit = 2 bytes per sample
+    let file_size = 36 + data_size;
+
+    let mut f = File::create(path)?;
+
+    // RIFF header
+    f.write_all(b"RIFF")?;
+    f.write_all(&file_size.to_le_bytes())?;
+    f.write_all(b"WAVE")?;
+
+    // fmt chunk — standard 16-byte PCM format
+    f.write_all(b"fmt ")?;
+    f.write_all(&16u32.to_le_bytes())?; // chunk size
+    f.write_all(&1u16.to_le_bytes())?; // audio format = PCM
+    f.write_all(&1u16.to_le_bytes())?; // channels = mono
+    f.write_all(&sample_rate.to_le_bytes())?;
+    f.write_all(&(sample_rate * 2).to_le_bytes())?; // byte rate
+    f.write_all(&2u16.to_le_bytes())?; // block align
+    f.write_all(&16u16.to_le_bytes())?; // bits per sample
+
+    // data chunk
+    f.write_all(b"data")?;
+    f.write_all(&data_size.to_le_bytes())?;
+
+    for &s in samples {
+        let clamped = s.clamp(-1.0, 1.0);
+        let pcm = (clamped * 32767.0) as i16;
+        f.write_all(&pcm.to_le_bytes())?;
+    }
+
+    Ok(())
+}
+
+pub struct KokorosService {
+    tts: Arc<TokioMutex<Option<TTSKoko>>>,
+    language: Arc<Mutex<String>>,
+    speed: Arc<Mutex<f32>>,
+}
+
+impl Default for KokorosService {
+    fn default() -> Self {
+        Self {
+            tts: Arc::new(TokioMutex::new(None)),
+            language: Arc::new(Mutex::new("en-us".to_string())),
+            speed: Arc::new(Mutex::new(1.0)),
+        }
+    }
+}
+
+#[tonic::async_trait]
+impl Backend for KokorosService {
+    async fn health(
+        &self,
+        _req: Request<backend::HealthMessage>,
+    ) -> Result<Response<backend::Reply>, Status> {
+        Ok(Response::new(backend::Reply {
+            message: b"OK".to_vec(),
+            ..Default::default()
+        }))
+    }
+
+    async fn load_model(
+        &self,
+        req: Request<backend::ModelOptions>,
+    ) -> Result<Response<backend::Result>, Status> {
+        let opts = req.into_inner();
+
+        // Model path: join ModelPath + Model, or just Model
+        let model_path = if !opts.model_path.is_empty() && !opts.model.is_empty() {
+            format!("{}/{}", opts.model_path, opts.model)
+        } else if !opts.model.is_empty() {
+            opts.model.clone()
+        } else {
+            "checkpoints/kokoro-v1.0.onnx".to_string()
+        };
+
+        // Voices data path from AudioPath, or derive from model dir
+        let voices_path = if !opts.audio_path.is_empty() {
+            opts.audio_path.clone()
+        } else {
+            let model_dir = std::path::Path::new(&model_path)
+                .parent()
+                .map(|p| p.to_string_lossy().to_string())
+                .unwrap_or_else(|| ".".to_string());
+            format!("{}/voices-v1.0.bin", model_dir)
+        };
+
+        // Parse options (key:value pairs)
+        for opt in &opts.options {
+            if let Some((key, value)) = opt.split_once(':') {
+                match key {
+                    "lang_code" => *self.language.lock().unwrap() = value.to_string(),
+                    "speed" => {
+                        if let Ok(s) = value.parse::<f32>() {
+                            *self.speed.lock().unwrap() = s;
+                        }
+                    }
+                    _ => {}
+                }
+            }
+        }
+
+        tracing::info!("Loading Kokoros model from: {}", model_path);
+        tracing::info!("Loading voices from: {}", voices_path);
+        tracing::info!("Language: {}", self.language.lock().unwrap());
+
+        let tts = TTSKoko::new(&model_path, &voices_path).await;
+        *self.tts.lock().await = Some(tts);
+
+        tracing::info!("Kokoros TTS model loaded successfully");
+        Ok(Response::new(backend::Result {
+            success: true,
+            message: "Kokoros TTS model loaded".into(),
+        }))
+    }
+
+    async fn tts(
+        &self,
+        req: Request<backend::TtsRequest>,
+    ) -> Result<Response<backend::Result>, Status> {
+        let req = req.into_inner();
+        let tts_guard = self.tts.lock().await;
+        let tts = tts_guard
+            .as_ref()
+            .ok_or_else(|| Status::failed_precondition("Model not loaded"))?;
+
+        let voice = if req.voice.is_empty() {
+            "af_heart"
+        } else {
+            &req.voice
+        };
+        let lang = req
+            .language
+            .filter(|l| !l.is_empty())
+            .unwrap_or_else(|| self.language.lock().unwrap().clone());
+        let speed = *self.speed.lock().unwrap();
+
+        tracing::info!(
+            text = req.text,
+            voice = voice,
+            lang = lang.as_str(),
+            dst = req.dst,
+            "TTS request received"
+        );
+
+        let start = std::time::Instant::now();
+        match tts.tts_raw_audio(&req.text, &lang, voice, speed, None, None, None, None) {
+            Ok(samples) => {
+                let duration_secs = samples.len() as f64 / 24000.0;
+                tracing::info!(
+                    num_samples = samples.len(),
+                    audio_duration = format!("{:.2}s", duration_secs),
+                    inference_time = format!("{:.2}s", start.elapsed().as_secs_f64()),
+                    dst = req.dst,
+                    "TTS inference complete"
+                );
+                if let Err(e) = write_pcm16_wav(&req.dst, &samples, 24000) {
+                    tracing::error!("Failed to write WAV to {}: {}", req.dst, e);
+                    return Ok(Response::new(backend::Result {
+                        success: false,
+                        message: format!("Failed to write WAV: {}", e),
+                    }));
+                }
+                Ok(Response::new(backend::Result {
+                    success: true,
+                    message: String::new(),
+                }))
+            }
+            Err(e) => {
+                tracing::error!("TTS error: {}", e);
+                Ok(Response::new(backend::Result {
+                    success: false,
+                    message: format!("TTS error: {}", e),
+                }))
+            }
+        }
+    }
+
+    type TTSStreamStream = ReceiverStream<Result<backend::Reply, Status>>;
+
+    async fn tts_stream(
+        &self,
+        req: Request<backend::TtsRequest>,
+    ) -> Result<Response<Self::TTSStreamStream>, Status> {
+        let req = req.into_inner();
+        let tts_guard = self.tts.lock().await;
+        let tts = tts_guard
+            .as_ref()
+            .ok_or_else(|| Status::failed_precondition("Model not loaded"))?
+            .clone();
+
+        let voice = if req.voice.is_empty() {
+            "af_heart".to_string()
+        } else {
+            req.voice
+        };
+        let lang = req
+            .language
+            .filter(|l| !l.is_empty())
+            .unwrap_or_else(|| self.language.lock().unwrap().clone());
+        let speed = *self.speed.lock().unwrap();
+        let text = req.text;
+
+        let (tx, rx) = tokio::sync::mpsc::channel(32);
+
+        // Send sample rate info as first message
+        let tx_clone = tx.clone();
+        let _ = tx_clone
+            .send(Ok(backend::Reply {
+                message: br#"{"sample_rate":24000}"#.to_vec(),
+                ..Default::default()
+            }))
+            .await;
+
+        tokio::task::spawn_blocking(move || {
+            let result = tts.tts_raw_audio_streaming(
+                &text,
+                &lang,
+                &voice,
+                speed,
+                None,
+                None,
+                None,
+                None,
+                |audio_chunk: Vec<f32>| -> Result<(), Box<dyn std::error::Error>> {
+                    // Convert f32 PCM to 16-bit PCM bytes (what LocalAI expects for streaming)
+                    let bytes: Vec<u8> = audio_chunk
+                        .iter()
+                        .flat_map(|&s| {
+                            let clamped = s.clamp(-1.0, 1.0);
+                            let i16_val = (clamped * 32767.0) as i16;
+                            i16_val.to_le_bytes()
+                        })
+                        .collect();
+                    tx.blocking_send(Ok(backend::Reply {
+                        audio: bytes,
+                        ..Default::default()
+                    }))
+                    .map_err(|e| Box::new(e) as Box<dyn std::error::Error>)
+                },
+            );
+            if let Err(e) = result {
+                tracing::error!("TTSStream error: {}", e);
+            }
+        });
+
+        Ok(Response::new(ReceiverStream::new(rx)))
+    }
+
+    async fn status(
+        &self,
+        _req: Request<backend::HealthMessage>,
+    ) -> Result<Response<backend::StatusResponse>, Status> {
+        let tts = self.tts.lock().await;
+        let state = if tts.is_some() {
+            backend::status_response::State::Ready as i32
+        } else {
+            backend::status_response::State::Uninitialized as i32
+        };
+        Ok(Response::new(backend::StatusResponse {
+            state,
+            memory: None,
+        }))
+    }
+
+    async fn free(
+        &self,
+        _req: Request<backend::HealthMessage>,
+    ) -> Result<Response<backend::Result>, Status> {
+        *self.tts.lock().await = None;
+        Ok(Response::new(backend::Result {
+            success: true,
+            message: "Model freed".into(),
+        }))
+    }
+
+    // --- Unimplemented RPCs ---
+
+    async fn predict(
+        &self,
+        _: Request<backend::PredictOptions>,
+    ) -> Result<Response<backend::Reply>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    type PredictStreamStream = ReceiverStream<Result<backend::Reply, Status>>;
+
+    async fn predict_stream(
+        &self,
+        _: Request<backend::PredictOptions>,
+    ) -> Result<Response<Self::PredictStreamStream>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn embedding(
+        &self,
+        _: Request<backend::PredictOptions>,
+    ) -> Result<Response<backend::EmbeddingResult>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn generate_image(
+        &self,
+        _: Request<backend::GenerateImageRequest>,
+    ) -> Result<Response<backend::Result>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn generate_video(
+        &self,
+        _: Request<backend::GenerateVideoRequest>,
+    ) -> Result<Response<backend::Result>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn audio_transcription(
+        &self,
+        _: Request<backend::TranscriptRequest>,
+    ) -> Result<Response<backend::TranscriptResult>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn sound_generation(
+        &self,
+        _: Request<backend::SoundGenerationRequest>,
+    ) -> Result<Response<backend::Result>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn tokenize_string(
+        &self,
+        _: Request<backend::PredictOptions>,
+    ) -> Result<Response<backend::TokenizationResponse>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn detect(
+        &self,
+        _: Request<backend::DetectOptions>,
+    ) -> Result<Response<backend::DetectResponse>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn stores_set(
+        &self,
+        _: Request<backend::StoresSetOptions>,
+    ) -> Result<Response<backend::Result>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn stores_delete(
+        &self,
+        _: Request<backend::StoresDeleteOptions>,
+    ) -> Result<Response<backend::Result>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn stores_get(
+        &self,
+        _: Request<backend::StoresGetOptions>,
+    ) -> Result<Response<backend::StoresGetResult>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn stores_find(
+        &self,
+        _: Request<backend::StoresFindOptions>,
+    ) -> Result<Response<backend::StoresFindResult>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn rerank(
+        &self,
+        _: Request<backend::RerankRequest>,
+    ) -> Result<Response<backend::RerankResult>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn get_metrics(
+        &self,
+        _: Request<backend::MetricsRequest>,
+    ) -> Result<Response<backend::MetricsResponse>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn vad(
+        &self,
+        _: Request<backend::VadRequest>,
+    ) -> Result<Response<backend::VadResponse>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn audio_encode(
+        &self,
+        _: Request<backend::AudioEncodeRequest>,
+    ) -> Result<Response<backend::AudioEncodeResult>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn audio_decode(
+        &self,
+        _: Request<backend::AudioDecodeRequest>,
+    ) -> Result<Response<backend::AudioDecodeResult>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn model_metadata(
+        &self,
+        _: Request<backend::ModelOptions>,
+    ) -> Result<Response<backend::ModelMetadataResponse>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn start_fine_tune(
+        &self,
+        _: Request<backend::FineTuneRequest>,
+    ) -> Result<Response<backend::FineTuneJobResult>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    type FineTuneProgressStream = ReceiverStream<Result<backend::FineTuneProgressUpdate, Status>>;
+
+    async fn fine_tune_progress(
+        &self,
+        _: Request<backend::FineTuneProgressRequest>,
+    ) -> Result<Response<Self::FineTuneProgressStream>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn stop_fine_tune(
+        &self,
+        _: Request<backend::FineTuneStopRequest>,
+    ) -> Result<Response<backend::Result>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn list_checkpoints(
+        &self,
+        _: Request<backend::ListCheckpointsRequest>,
+    ) -> Result<Response<backend::ListCheckpointsResponse>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn export_model(
+        &self,
+        _: Request<backend::ExportModelRequest>,
+    ) -> Result<Response<backend::Result>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn start_quantization(
+        &self,
+        _: Request<backend::QuantizationRequest>,
+    ) -> Result<Response<backend::QuantizationJobResult>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    type QuantizationProgressStream =
+        ReceiverStream<Result<backend::QuantizationProgressUpdate, Status>>;
+
+    async fn quantization_progress(
+        &self,
+        _: Request<backend::QuantizationProgressRequest>,
+    ) -> Result<Response<Self::QuantizationProgressStream>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn stop_quantization(
+        &self,
+        _: Request<backend::QuantizationStopRequest>,
+    ) -> Result<Response<backend::Result>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn wav_header_is_standard_pcm16() {
+        let samples = vec![0.0f32, 0.5, -0.5, 1.0, -1.0];
+        let path = std::env::temp_dir().join("kokoros_test.wav");
+        let path_str = path.to_str().unwrap();
+
+        write_pcm16_wav(path_str, &samples, 24000).unwrap();
+
+        let data = std::fs::read(&path).unwrap();
+        std::fs::remove_file(&path).unwrap();
+
+        // Must be exactly 44-byte header + data
+        assert_eq!(data.len(), 44 + samples.len() * 2);
+
+        // RIFF header
+        assert_eq!(&data[0..4], b"RIFF");
+        assert_eq!(&data[8..12], b"WAVE");
+
+        // fmt chunk: 16 bytes, format=1 (PCM), channels=1, 16-bit
+        assert_eq!(&data[12..16], b"fmt ");
+        assert_eq!(u32::from_le_bytes(data[16..20].try_into().unwrap()), 16); // chunk size
+        assert_eq!(u16::from_le_bytes(data[20..22].try_into().unwrap()), 1); // PCM format
+        assert_eq!(u16::from_le_bytes(data[22..24].try_into().unwrap()), 1); // mono
+        assert_eq!(u32::from_le_bytes(data[24..28].try_into().unwrap()), 24000); // sample rate
+        assert_eq!(u16::from_le_bytes(data[34..36].try_into().unwrap()), 16); // bits per sample
+
+        // data chunk
+        assert_eq!(&data[36..40], b"data");
+        assert_eq!(
+            u32::from_le_bytes(data[40..44].try_into().unwrap()),
+            (samples.len() * 2) as u32
+        );
+
+        // Verify sample values: 0.5 -> 16383, -0.5 -> -16383, 1.0 -> 32767, -1.0 -> -32767
+        let s1 = i16::from_le_bytes(data[46..48].try_into().unwrap());
+        assert_eq!(s1, 16383); // 0.5 * 32767
+        let s3 = i16::from_le_bytes(data[50..52].try_into().unwrap());
+        assert_eq!(s3, 32767); // 1.0 clamped
+        let s4 = i16::from_le_bytes(data[52..54].try_into().unwrap());
+        assert_eq!(s4, -32767); // -1.0 clamped
+    }
+
+    /// Integration test: runs actual TTS inference and validates the output audio.
+    /// Skipped unless KOKOROS_MODEL_PATH is set to a directory containing
+    /// kokoro-v1.0.onnx and voices-v1.0.bin.
+    #[tokio::test]
+    async fn tts_produces_valid_speech() {
+        let model_dir = match std::env::var("KOKOROS_MODEL_PATH") {
+            Ok(p) => p,
+            Err(_) => {
+                eprintln!("KOKOROS_MODEL_PATH not set, skipping integration test");
+                return;
+            }
+        };
+
+        let model_path = format!("{}/kokoro-v1.0.onnx", model_dir);
+        let voices_path = format!("{}/voices-v1.0.bin", model_dir);
+
+        if !std::path::Path::new(&model_path).exists() {
+            eprintln!("Model file not found at {}, skipping", model_path);
+            return;
+        }
+
+        let tts = TTSKoko::new(&model_path, &voices_path).await;
+
+        let input_text = "Hello world, this is a test of speech synthesis.";
+        let out_path = std::env::temp_dir().join("kokoros_integration_test.wav");
+        let out_str = out_path.to_str().unwrap();
+
+        let samples = tts
+            .tts_raw_audio(input_text, "en-us", "af_heart", 1.0, None, None, None, None)
+            .expect("tts_raw_audio failed");
+
+        write_pcm16_wav(out_str, &samples, 24000).unwrap();
+
+        let data = std::fs::read(&out_path).unwrap();
+        std::fs::remove_file(&out_path).unwrap();
+
+        // --- WAV header sanity ---
+        assert_eq!(&data[0..4], b"RIFF");
+        assert_eq!(&data[8..12], b"WAVE");
+        assert_eq!(u16::from_le_bytes(data[20..22].try_into().unwrap()), 1); // PCM
+        assert_eq!(u32::from_le_bytes(data[24..28].try_into().unwrap()), 24000); // sample rate
+        assert_eq!(u16::from_le_bytes(data[34..36].try_into().unwrap()), 16); // 16-bit
+
+        let num_samples = samples.len();
+        let duration_secs = num_samples as f64 / 24000.0;
+
+        // --- Duration check ---
+        // ~10 words should produce roughly 2-8 seconds of speech
+        assert!(
+            duration_secs > 1.0,
+            "Audio too short: {:.2}s for {} words",
+            duration_secs,
+            input_text.split_whitespace().count()
+        );
+        assert!(
+            duration_secs < 15.0,
+            "Audio too long: {:.2}s for {} words",
+            duration_secs,
+            input_text.split_whitespace().count()
+        );
+
+        // --- Energy check: not silence ---
+        let rms = (samples.iter().map(|s| s * s).sum::<f32>() / num_samples as f32).sqrt();
+        assert!(
+            rms > 0.01,
+            "Audio is near-silence: RMS = {:.6}",
+            rms
+        );
+
+        // --- Not clipped/saturated: should have dynamic range ---
+        let max_abs = samples.iter().map(|s| s.abs()).fold(0.0f32, f32::max);
+        assert!(
+            max_abs < 1.0,
+            "Audio is fully saturated (max |sample| = {:.4})",
+            max_abs
+        );
+        assert!(
+            max_abs > 0.05,
+            "Audio has very low amplitude (max |sample| = {:.4})",
+            max_abs
+        );
+
+        // --- Speech-like spectral check ---
+        // Speech should have significant energy variation (not white noise or DC).
+        // Check that the signal has zero-crossings in a speech-like range (roughly
+        // 50-400 crossings per 24000 samples = 100-8000 Hz fundamental range).
+        let zero_crossings: usize = samples
+            .windows(2)
+            .filter(|w| (w[0] >= 0.0) != (w[1] >= 0.0))
+            .count();
+        let crossings_per_sec = zero_crossings as f64 / duration_secs;
+        // White noise at 24kHz would have ~12000 crossings/sec.
+        // Speech is typically 100-4000 crossings/sec.
+        assert!(
+            crossings_per_sec < 10000.0,
+            "Too many zero crossings ({:.0}/s) — likely noise, not speech",
+            crossings_per_sec
+        );
+        assert!(
+            crossings_per_sec > 50.0,
+            "Too few zero crossings ({:.0}/s) — likely DC or silence, not speech",
+            crossings_per_sec
+        );
+
+        eprintln!(
+            "Integration test passed: duration={:.2}s, rms={:.4}, max={:.4}, zero_crossings={:.0}/s",
+            duration_secs, rms, max_abs, crossings_per_sec
+        );
+    }
+}

gallery/index.yaml

@@ -2514,6 +2514,132 @@
       - lang_code:a
     known_usecases:
       - tts
+- name: "kokoros"
+  url: "github:mudler/LocalAI/gallery/kokoros.yaml@master"
+  size: "327MB"
+  urls:
+    - https://github.com/lucasjinreal/Kokoros
+  license: apache-2.0
+  tags:
+    - tts
+    - kokoros
+    - cpu
+    - text-to-speech
+    - rust
+  description: |
+    Kokoros is a pure Rust TTS backend using the Kokoro v1.0 ONNX model (82M parameters).
+    Fast, streaming TTS with high quality. American English with af_heart voice.
+  overrides:
+    backend: "kokoros"
+    name: "kokoros"
+    description: "Kokoros Rust TTS - American English"
+    parameters:
+      model: "kokoro-v1.0.onnx"
+      voice: "af_heart"
+    options:
+      - lang_code:en-us
+    known_usecases:
+      - tts
+  files:
+    - filename: "kokoro-v1.0.onnx"
+      sha256: 7d5df8ecf7d4b1878015a32686053fd0eebe2bc377234608764cc0ef3636a6c5
+      uri: "https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.onnx"
+    - filename: "voices-v1.0.bin"
+      sha256: bca610b8308e8d99f32e6fe4197e7ec01679264efed0cac9140fe9c29f1fbf7d
+      uri: "https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/voices-v1.0.bin"
+- name: "kokoros-ja"
+  url: "github:mudler/LocalAI/gallery/kokoros.yaml@master"
+  size: "327MB"
+  urls:
+    - https://github.com/lucasjinreal/Kokoros
+  license: apache-2.0
+  tags:
+    - tts
+    - kokoros
+    - japanese
+    - text-to-speech
+  description: |
+    Kokoros Rust TTS - Japanese. Uses the Kokoro v1.0 ONNX model with Japanese phonemization.
+  overrides:
+    backend: "kokoros"
+    name: "kokoros-ja"
+    description: "Kokoros Rust TTS - Japanese"
+    parameters:
+      model: "kokoro-v1.0.onnx"
+      voice: "jf_alpha"
+    options:
+      - lang_code:ja
+    known_usecases:
+      - tts
+  files:
+    - filename: "kokoro-v1.0.onnx"
+      sha256: 7d5df8ecf7d4b1878015a32686053fd0eebe2bc377234608764cc0ef3636a6c5
+      uri: "https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.onnx"
+    - filename: "voices-v1.0.bin"
+      sha256: bca610b8308e8d99f32e6fe4197e7ec01679264efed0cac9140fe9c29f1fbf7d
+      uri: "https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/voices-v1.0.bin"
+- name: "kokoros-cmn"
+  url: "github:mudler/LocalAI/gallery/kokoros.yaml@master"
+  size: "327MB"
+  urls:
+    - https://github.com/lucasjinreal/Kokoros
+  license: apache-2.0
+  tags:
+    - tts
+    - kokoros
+    - chinese
+    - text-to-speech
+  description: |
+    Kokoros Rust TTS - Mandarin Chinese.
+  overrides:
+    backend: "kokoros"
+    name: "kokoros-cmn"
+    description: "Kokoros Rust TTS - Mandarin Chinese"
+    parameters:
+      model: "kokoro-v1.0.onnx"
+      voice: "zf_xiaobei"
+    options:
+      - lang_code:cmn
+    known_usecases:
+      - tts
+  files:
+    - filename: "kokoro-v1.0.onnx"
+      sha256: 7d5df8ecf7d4b1878015a32686053fd0eebe2bc377234608764cc0ef3636a6c5
+      uri: "https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.onnx"
+    - filename: "voices-v1.0.bin"
+      sha256: bca610b8308e8d99f32e6fe4197e7ec01679264efed0cac9140fe9c29f1fbf7d
+      uri: "https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/voices-v1.0.bin"
+- name: "kokoros-de"
+  url: "github:mudler/LocalAI/gallery/kokoros.yaml@master"
+  size: "327MB"
+  urls:
+    - https://github.com/lucasjinreal/Kokoros
+  license: apache-2.0
+  tags:
+    - tts
+    - kokoros
+    - german
+    - text-to-speech
+  description: |
+    Kokoros Rust TTS - German.
+  overrides:
+    backend: "kokoros"
+    name: "kokoros-de"
+    description: "Kokoros Rust TTS - German"
+    parameters:
+      model: "kokoro-v1.0.onnx"
+      voice: "df_greta"
+    options:
+      - lang_code:de
+    known_usecases:
+      - tts
+  files:
+    - filename: "kokoro-v1.0.onnx"
+      sha256: 7d5df8ecf7d4b1878015a32686053fd0eebe2bc377234608764cc0ef3636a6c5
+      uri: "https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.onnx"
+    - filename: "voices-v1.0.bin"
+      sha256: bca610b8308e8d99f32e6fe4197e7ec01679264efed0cac9140fe9c29f1fbf7d
+      uri: "https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/voices-v1.0.bin"
 - name: "kitten-tts"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:

gallery/kokoros.yaml

@@ -0,0 +1,3 @@
+---
+config_file: |
+  backend: kokoros

scripts/changed-backends.js

@@ -21,6 +21,9 @@ function inferBackendPath(item) {
   if (item.dockerfile.endsWith("golang")) {
     return `backend/go/${item.backend}/`;
   }
+  if (item.dockerfile.endsWith("rust")) {
+    return `backend/rust/${item.backend}/`;
+  }
   if (item.dockerfile.endsWith("llama-cpp")) {
     return `backend/cpp/llama-cpp/`;
   }