Nik Afiq 5238298b55
Some checks failed
CI / test (push) Successful in 6s
CI / build-ai-gateway (push) Failing after 28s
CI / build-ha-gateway (push) Failing after 24s
CI / build-discord-bot (push) Failing after 29s
Add TTS model components and inference server
- Implemented core model components in `modules.py` including various convolutional layers and normalization techniques.
- Added transformation functions in `transforms.py` for piecewise rational quadratic transformations.
- Created utility functions in `utils.py` for checkpoint management, logging, and hyperparameter handling.
- Introduced monotonic alignment functionality with Cython optimization in `monotonic_align`.
- Developed a minimal inference server in `server.py` to handle synthesis requests.
- Updated requirements to include necessary dependencies for Cython and scipy.
2026-07-24 22:38:36 +09:00

97 lines
3.0 KiB
Go

// Package inferencesidecar implements driven.TTSEngine by calling a small
// Python/libtorch sidecar (tts-gateway/sidecar) over HTTP - the fallback
// documented in TTS_GATEWAY_PLAN.md's Risk 1 and adopted after the Phase 0
// spike found the checkpoint's data-dependent output length isn't cleanly
// exportable via torch.export. Same shape as ai-gateway's Ollama client:
// the tensor math lives outside Go, hidden behind this port like any other
// outbound dependency.
package inferencesidecar
import (
"bytes"
"context"
"encoding/binary"
"encoding/json"
"fmt"
"io"
"math"
"net/http"
"strconv"
"time"
"gitea.nik4nao.com/nik/home-services/tts-gateway/internal/core/domain"
)
type Client struct {
baseURL string
httpClient *http.Client
}
// NewClient constructs the sidecar HTTP client. addr is a host:port, e.g.
// "localhost:50054" (config.Config.InferenceSidecarAddr).
func NewClient(addr string) *Client {
return &Client{
baseURL: "http://" + addr,
httpClient: &http.Client{Timeout: 60 * time.Second},
}
}
type synthesizeRequest struct {
SpeakerID int `json:"speaker_id"`
SymbolIDs []int64 `json:"symbol_ids"`
NoiseScale float32 `json:"noise_scale"`
NoiseScaleW float32 `json:"noise_scale_w"`
LengthScale float32 `json:"length_scale"`
}
// Synthesize posts symbol IDs + params to the sidecar's /synthesize
// endpoint and decodes its raw little-endian float32 PCM response.
func (c *Client) Synthesize(ctx context.Context, speakerID int, symbolIDs []int64, params domain.SynthesisParams) ([]float32, int, error) {
reqBody, err := json.Marshal(synthesizeRequest{
SpeakerID: speakerID,
SymbolIDs: symbolIDs,
NoiseScale: params.NoiseScale,
NoiseScaleW: params.NoiseScaleW,
LengthScale: params.LengthScale,
})
if err != nil {
return nil, 0, fmt.Errorf("inferencesidecar: marshal request: %w", err)
}
req, err := http.NewRequestWithContext(ctx, http.MethodPost, c.baseURL+"/synthesize", bytes.NewReader(reqBody))
if err != nil {
return nil, 0, fmt.Errorf("inferencesidecar: build request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
resp, err := c.httpClient.Do(req)
if err != nil {
return nil, 0, fmt.Errorf("inferencesidecar: request failed: %w", err)
}
defer resp.Body.Close()
data, err := io.ReadAll(resp.Body)
if err != nil {
return nil, 0, fmt.Errorf("inferencesidecar: read response: %w", err)
}
if resp.StatusCode != http.StatusOK {
return nil, 0, fmt.Errorf("inferencesidecar: %s: %s", resp.Status, string(data))
}
sampleRate, err := strconv.Atoi(resp.Header.Get("X-Sample-Rate"))
if err != nil {
return nil, 0, fmt.Errorf("inferencesidecar: missing/invalid X-Sample-Rate header: %w", err)
}
if len(data)%4 != 0 {
return nil, 0, fmt.Errorf("inferencesidecar: response length %d not a multiple of 4", len(data))
}
pcm := make([]float32, len(data)/4)
for i := range pcm {
bits := binary.LittleEndian.Uint32(data[i*4 : i*4+4])
pcm[i] = math.Float32frombits(bits)
}
return pcm, sampleRate, nil
}