#!/usr/bin/env python3 """the engine Space: a streaming web UI around Fish Audio the engine on ZeroGPU. This process serves Gradio and runs the engine itself: zerogpu_backend loads the model at startup and streams each take from one @spaces.GPU call, through the same /v1/audio/speech request and events the SGLang Omni server used. Whisper transcribes reference clips on CPU. Voices come three ways: cloned from a clip, cast from a text description, or left to the model. A take can be redone with a director's note; the earlier takes ride along in the prompt so the note is read against what the listener heard (engine_engine.py builds those prompts). There is no daily take allowance: ZeroGPU applies each visitor's own GPU quota to their takes. Per-account concurrency and burst limits keep one caller from filling the queue (accounts.py identifies callers). """ from __future__ import annotations import asyncio import base64 import collections import contextlib import hashlib import io import json import math import os import re import secrets import sys import threading import time from dataclasses import dataclass, field, replace from pathlib import Path from typing import Literal import httpx import numpy as np import soundfile as sf import uvicorn from fastapi import Depends, FastAPI, File, HTTPException, Request, UploadFile from fastapi.responses import HTMLResponse, Response, StreamingResponse from pydantic import BaseModel, Field from stream_transport import StreamingGZipMiddleware from accounts import ( Identity, identify, public_user, ) APP_DIR = Path(__file__).resolve().parent VOICES_DIR = APP_DIR / "voices" def env_int(name: str, default: int) -> int: return int(os.environ.get(name, default)) PORT = env_int("PORT", 7860) # The in-process engine takes the speech server's request body; the URL is a label. ENGINE_URL = "zerogpu://engine" # Takes this server runs at once (each on its own ZeroGPU GPU); the rest wait in a # queue that reports position. MAX_ACTIVE = env_int("DEMO_MAX_ACTIVE", 12) MAX_QUEUE = env_int("DEMO_MAX_QUEUE", 64) # Per account: concurrent takes, and a burst cap. GPU time itself is bounded by each # visitor's ZeroGPU quota. PER_CLIENT_ACTIVE = env_int("DEMO_PER_CLIENT_ACTIVE", 2) PER_CLIENT_REQUESTS = env_int("DEMO_PER_CLIENT_REQUESTS", 60) RATE_WINDOW_S = env_int("DEMO_RATE_WINDOW_S", 600) MAX_TEXT_CHARS = 800 # ~54 s of speech, just under the 1200-frame (55.7 s) take cap MAX_TRANSCRIPT_CHARS = 600 MAX_NOTE_CHARS = 120 MAX_NAME_CHARS = 40 MAX_DESCRIPTION_CHARS = 300 MAX_SCENE_CHARS = 200 # the engine frames are 2048 samples at 44.1 kHz, about 21.5 per second: 1200 frames is # about 56 s, under the engine's 60 s limit for a take that returns as a reference. MAX_NEW_TOKENS = 1200 # the engine falls silent mid-take under top-k truncation: at a pause a few silence # codes outrank the many ways speech can resume, and top-k 30 keeps only the former. # On an H200, top-k 30 (the SGLang Omni default) left 11 of 40 takes silent to the # length cap, and no cap left none. Requests send this many candidates instead: the # end-of-speech token and all 4,096 semantic codes, which also covers the 4,096-entry # codebook heads. A top_k of 0 asks for it. ALL_CANDIDATES = 4097 MAX_NEW_TOKENS_CAP = 1200 # matches engine.py: the frame cap of one take SAMPLE_RATE = 44100 REF_MAX_SECONDS = 20.0 REF_MIN_SECONDS = 1.0 UPLOAD_MAX_BYTES = 25 * 1024 * 1024 # Earlier takes kept in the prompt of a retake. HISTORY_TAKES = 2 # Keep short natural pauses, trim leading/trailing silence, and abort sustained # silence. Keep the existing threshold so quiet speech is still passed through. SILENCE_RMS = 0.003 MAX_PAUSE_S = 0.75 TAIL_S = 0.15 # Hold only the beginning so short failed fragments can be retried before playback. # The first four engine windows total 64 frames (2.972 s). A 3 s threshold # misses that boundary and waits for the next 64-frame window. Keep room for # trimming quiet edges while still withholding short failed openings. START_BUFFER_S = 2.5 # A whole opening buffer is a large SSE line; clients must receive that entire # line before playing anything. Send half-second packets after validating it. AUDIO_PACKET_S = 0.5 STALL_LIMIT_S = 4.0 KEEPALIVE_S = 5.0 # Until the first audio: a ZeroGPU queue, a fresh GPU worker (weights to the GPU, # compile and CUDA-graph capture) and the prefill. ENGINE_START_TIMEOUT_S = float(os.environ.get("ENGINE_START_TIMEOUT_S", 150)) ENGINE_IDLE_TIMEOUT_S = 15.0 TAKE_TIMEOUT_S = float(os.environ.get("TAKE_TIMEOUT_S", 180)) TAKE_STORE_BYTES = 512 * 1024 * 1024 VOICE_STORE_SIZE = 256 WHISPER_MODEL = os.environ.get("WHISPER_MODEL", "turbo") WHISPER_DIR = os.environ.get("WHISPER_DIR", str(Path.home() / ".cache" / "whisper")) # Must match engine_engine.ENVELOPE_MARKER (kept separate so the UI imports no torch). ENVELOPE_MARKER = "@@tts-envelope@@" STRUCTURAL = re.compile(r"<\|") DIRECTION_SPAN = re.compile(r"\[[^\[\]\n]*\]|\n]{0,40}>") def log(message: str) -> None: print(f"[space] {message}", flush=True) # ─── Engine ─────────────────────────────────────────────────────────────────── class Engine: """Status of the in-process engine (zerogpu_backend), for the boot screen.""" def __init__(self) -> None: self.state: Literal["starting", "ready", "error"] = "starting" self.phase = "Loading 41 GB of weights" self.started_at = time.time() self.ready_at: float | None = None self.last_line = "" # The prompt layouts for described characters and directing history. self.plugin = False def use_in_process(self) -> None: """The weights loaded before the server started; GPUs attach per take.""" self.state = "ready" self.phase = "Ready" self.ready_at = time.time() self.plugin = True def status(self) -> dict: return { "state": self.state, "phase": self.phase, "elapsed_s": round(time.time() - self.started_at), "ready_s": round(self.ready_at - self.started_at) if self.ready_at else None, "log": self.last_line, "plugin": self.plugin, } # ─── Transcription ──────────────────────────────────────────────────────────── class Transcriber: """Whisper on CPU, so transcribing a clip never takes a visitor's GPU time.""" def __init__(self) -> None: self.model = None self.state: Literal["waiting", "loading", "ready", "error"] = "waiting" self.lock = threading.Lock() def start(self) -> None: if self.state == "waiting": self.state = "loading" threading.Thread(target=self.load, daemon=True).start() def load(self) -> None: try: from faster_whisper import WhisperModel self.model = WhisperModel(WHISPER_MODEL, device="cpu", compute_type="int8", cpu_threads=env_int("WHISPER_CPU_THREADS", 2), download_root=WHISPER_DIR) self.state = "ready" log(f"whisper {WHISPER_MODEL} ready") except Exception as error: # transcription is optional; typing still works self.state = "error" log(f"whisper failed to load: {error!r}") def transcribe(self, wav: np.ndarray, sample_rate: int) -> tuple[str, str | None]: import librosa audio = librosa.resample(wav, orig_sr=sample_rate, target_sr=16000).astype(np.float32) with self.lock: segments, info = self.model.transcribe( audio, beam_size=1, temperature=0.0, condition_on_previous_text=False ) return " ".join(segment.text.strip() for segment in segments).strip(), info.language # ─── Audio helpers ──────────────────────────────────────────────────────────── def decode_audio(data: bytes) -> tuple[np.ndarray, int]: """Mono float32 audio from any common container (WAV, MP3, M4A, OGG, WebM...).""" try: wav, sample_rate = sf.read(io.BytesIO(data), dtype="float32", always_2d=True) return wav.mean(axis=1), sample_rate except Exception: pass import av parts = [] with av.open(io.BytesIO(data)) as container: stream = container.streams.audio[0] resampler = av.AudioResampler(format="flt", layout="mono", rate=SAMPLE_RATE) for frame in container.decode(stream): parts += [out.to_ndarray().reshape(-1) for out in resampler.resample(frame)] parts += [out.to_ndarray().reshape(-1) for out in resampler.resample(None)] return (np.concatenate(parts) if parts else np.zeros(0, np.float32)), SAMPLE_RATE def tidy_reference(wav: np.ndarray, sample_rate: int) -> tuple[np.ndarray, bool]: """44.1 kHz, edge silence trimmed, at most REF_MAX_SECONDS cut at a quiet spot.""" import librosa if sample_rate != SAMPLE_RATE: wav = librosa.resample(wav, orig_sr=sample_rate, target_sr=SAMPLE_RATE) wav = np.nan_to_num(wav.astype(np.float32)) trimmed, _ = librosa.effects.trim(wav, top_db=40, frame_length=1024, hop_length=256) if trimmed.size: wav = trimmed cut = False limit = int(REF_MAX_SECONDS * SAMPLE_RATE) if wav.size > limit: # End at the quietest 150 ms in the last 40% of the window, so the cut lands # between words and the transcript of the kept part stays exact. start = int(0.6 * limit) window, hop = int(0.15 * SAMPLE_RATE), int(0.01 * SAMPLE_RATE) frames = np.lib.stride_tricks.sliding_window_view(wav[start:limit], window)[::hop] quietest = int(np.argmin(np.sqrt((frames**2).mean(axis=1)))) wav = wav[: start + quietest * hop + window // 2].copy() fade = min(int(0.03 * SAMPLE_RATE), wav.size) wav[-fade:] *= np.linspace(1.0, 0.0, fade, dtype=np.float32) cut = True peak = float(np.abs(wav).max()) if wav.size else 0.0 if peak > 0: wav = wav * min(10.0, 0.9 / peak) return wav.astype(np.float32), cut def wav_bytes(wav: np.ndarray, sample_rate: int) -> bytes: buffer = io.BytesIO() sf.write(buffer, np.clip(wav, -1.0, 1.0), sample_rate, format="WAV", subtype="PCM_16") return buffer.getvalue() def pcm_to_float(pcm: bytes) -> np.ndarray: return np.frombuffer(pcm, dtype=" bytes: """Lossless PCM16 references: the GPU process decodes the exact samples.""" buffer = io.BytesIO() sf.write(buffer, samples, sample_rate, format="FLAC", subtype="PCM_16") return buffer.getvalue() def spoken_text(line: str) -> str: """The words of a line, without direction: what a listener hears.""" match = re.search(r"^\s*Line:[ \t]*(.+)$", line, re.M | re.S) if match and re.match(r"\s*Direction:", line): line = match.group(1) return re.sub(r"\s+", " ", DIRECTION_SPAN.sub(" ", line)).strip() def token_budget(body: GenerateBody, choice: VoiceChoice) -> int: """A generous speech duration limit, with extra room for slow directions. This bounds runaway *audible* output as well as silence. It is a safety limit, not an estimate that the model necessarily completes the script within it. """ text = spoken_text(body.text) # Include languages without spaces; a word-count alone underbudgets them. cjk = len(re.findall(r"[\u3040-\u30ff\u3400-\u9fff\uac00-\ud7af]", text)) other = re.sub(r"[\u3040-\u30ff\u3400-\u9fff\uac00-\ud7af]", " ", text) units = sum(max(1, len(word) / 5) for word in other.split()) + cjk / 2 seconds = max(18.0, 10.0 + 1.2 * units) direction = body.text + " " + (body.note or "") if choice.character: direction += " " + choice.character.description + " " + (choice.character.scene or "") if re.search(r"\b(slow|slowly|drawn[- ]out|unhurried|long pause)\b", direction, re.I): seconds *= 1.6 return min(MAX_NEW_TOKENS, math.ceil(seconds * SAMPLE_RATE / 2048)) class SilenceGuard: """Buffer quiet frames so empty takes and trailing silence never go on air.""" def __init__(self, sample_rate: int) -> None: self.frame = max(1, int(0.05 * sample_rate)) self.max_pause = int(MAX_PAUSE_S * sample_rate) self.tail = int(TAIL_S * sample_rate) self.stall_limit = int(STALL_LIMIT_S * sample_rate) self.pending = b"" self.quiet = bytearray() self.silent_run = 0 self.dropped = 0 self.speech_detected = False self.stalled = False def feed(self, pcm: bytes, *, final: bool = False) -> bytes: data = self.pending + pcm samples = np.frombuffer(data[: len(data) // 2 * 2], dtype="= self.stall_limit: self.stalled = True else: self.speech_detected = True self.silent_run = 0 kept.extend(self.quiet) self.quiet.clear() kept.extend(frame.tobytes()) if final: hold = min(self.tail, len(self.quiet) // 2) kept.extend(self.quiet[:hold * 2]) self.dropped += len(self.quiet) // 2 - hold self.quiet.clear() return bytes(kept) async def engine_lines(response, started: float, progress: dict): """Bound time without audio even when the server keeps sending heartbeats.""" iterator = response.aiter_lines().__aiter__() pending = None try: while True: now = time.perf_counter() idle = ENGINE_IDLE_TIMEOUT_S if progress["has_audio"] else ENGINE_START_TIMEOUT_S remaining = min(TAKE_TIMEOUT_S - (now - started), idle - (now - progress["audio_at"])) if remaining <= 0: raise asyncio.TimeoutError if pending is None: pending = asyncio.ensure_future(anext(iterator)) if not (await asyncio.wait({pending}, timeout=min(KEEPALIVE_S, remaining)))[0]: yield None continue try: raw = pending.result() except StopAsyncIteration: break pending = None yield raw finally: if pending is not None: pending.cancel() await asyncio.gather(pending, return_exceptions=True) if hasattr(iterator, "aclose"): await iterator.aclose() # ─── Voices and takes ───────────────────────────────────────────────────────── @dataclass class Voice: id: str label: str wav: bytes transcript: str duration: float preset: bool = False language: str | None = None b64: str = field(init=False, repr=False) def __post_init__(self) -> None: # Keep WAV for browser downloads, and cache the lossless engine reference. samples, rate = sf.read(io.BytesIO(self.wav), dtype="int16") self.b64 = base64.b64encode(flac_bytes(samples, rate)).decode("ascii") def public(self) -> dict: return { "id": self.id, "label": self.label, "transcript": self.transcript, "duration": round(self.duration, 2), "language": self.language, "preset": self.preset, } @dataclass class Character: name: str description: str scene: str | None @dataclass class VoiceChoice: mode: Literal["clone", "design", "random"] voice: Voice | None = None transcript: str | None = None character: Character | None = None # A Surprise/Design voice becomes concrete after its first generated take. # Keep that reference across retakes without changing the visible voice mode. directing_voice: Voice | None = None @dataclass class Take: id: str choice: VoiceChoice line: str turn_text: str history: list[str] pcm: bytes sample_rate: int @property def seconds(self) -> float: return len(self.pcm) / 2 / self.sample_rate class Store: def __init__(self) -> None: self.voices: collections.OrderedDict[str, Voice] = collections.OrderedDict() self.presets: dict[str, Voice] = {} self.takes: collections.OrderedDict[str, Take] = collections.OrderedDict() self.take_bytes = 0 def load_presets(self) -> None: for entry in json.loads((VOICES_DIR / "voices.json").read_text()): # A clip ships either as a file or inline as base64, because binary # files need LFS/Xet on the Hub and a metadata-only commit cannot carry one. if entry.get("audio_b64"): raw = base64.b64decode(entry["audio_b64"]) else: path = VOICES_DIR / str(entry.get("file")) if not path.exists(): log(f"preset {entry['id']} skipped: {path.name} is missing") continue raw = path.read_bytes() wav, sample_rate = decode_audio(raw) wav, _ = tidy_reference(wav, sample_rate) self.presets[entry["id"]] = Voice( id=entry["id"], label=entry["label"], wav=wav_bytes(wav, SAMPLE_RATE), transcript=entry["transcript"], duration=wav.size / SAMPLE_RATE, preset=True, language="en", ) log(f"loaded {len(self.presets)} preset voices") def voice(self, voice_id: str) -> Voice | None: if voice_id in self.presets: return self.presets[voice_id] voice = self.voices.get(voice_id) if voice is not None: self.voices.move_to_end(voice_id) return voice def add_voice(self, voice: Voice) -> None: self.voices[voice.id] = voice self.voices.move_to_end(voice.id) while len(self.voices) > VOICE_STORE_SIZE: self.voices.popitem(last=False) def take(self, take_id: str) -> Take | None: return self.takes.get(take_id) def add_take(self, take: Take) -> None: self.takes[take.id] = take self.take_bytes += len(take.pcm) while self.take_bytes > TAKE_STORE_BYTES and len(self.takes) > 1: _, old = self.takes.popitem(last=False) self.take_bytes -= len(old.pcm) # ─── Admission ──────────────────────────────────────────────────────────────── class Ticket: def __init__(self) -> None: self.granted = asyncio.Event() class Gate: """FIFO admission to MAX_ACTIVE engine streams; waiters learn their position.""" def __init__(self, slots: int) -> None: self.slots = slots self.active = 0 self.waiting: collections.deque[Ticket] = collections.deque() def enqueue(self) -> Ticket: ticket = Ticket() if self.active < self.slots and not self.waiting: self.active += 1 ticket.granted.set() else: self.waiting.append(ticket) return ticket def position(self, ticket: Ticket) -> int: try: return self.waiting.index(ticket) + 1 except ValueError: return 0 def leave(self, ticket: Ticket) -> None: if ticket.granted.is_set(): self.active -= 1 else: self.waiting.remove(ticket) while self.waiting and self.active < self.slots: self.active += 1 self.waiting.popleft().granted.set() class ClientLimits: """Per-account concurrency and burst limits. Expiration is a final recovery guard. Streams reserve their slot only once consumption starts, and release it when their iterator closes. """ SLOT_TTL_S = 300 def __init__(self) -> None: self.active: dict[str, list[float]] = collections.defaultdict(list) self.recent: dict[str, collections.deque[float]] = collections.defaultdict(collections.deque) def check(self, client: str) -> str | None: """Check limits without reserving a slot or recording a new request.""" now = time.time() recent = self.recent[client] while recent and now - recent[0] > RATE_WINDOW_S: recent.popleft() slots = [t for t in self.active.get(client, []) if now - t < self.SLOT_TTL_S] if slots: self.active[client] = slots else: self.active.pop(client, None) if len(slots) >= PER_CLIENT_ACTIVE: return f"You already have {PER_CLIENT_ACTIVE} takes in progress. Wait for one to finish." if len(recent) >= PER_CLIENT_REQUESTS: wait = int(RATE_WINDOW_S - (now - recent[0])) + 1 return f"That's a lot of takes! Give the GPU a breather and try again in {wait}s." return None def admit(self, client: str) -> float | str: """Reserve atomically when a response actually begins streaming.""" refused = self.check(client) if refused: return refused now = time.time() self.recent[client].append(now) self.active[client].append(now) return now def release(self, client: str, slot: float) -> None: slots = self.active.get(client, []) if slot in slots: slots.remove(slot) if not slots: self.active.pop(client, None) # ─── App state ──────────────────────────────────────────────────────────────── engine = Engine() transcriber = Transcriber() store = Store() gate = Gate(MAX_ACTIVE) limits = ClientLimits() stats = {"takes": 0, "audio_s": 0.0} recent_ttfa: collections.deque[float] = collections.deque(maxlen=50) http = None # zerogpu_backend.EngineClient once the model is loaded def gpu_info() -> dict | None: # GPUs attach per take, so this process has no device to report on. return {"name": "ZeroGPU", "remote": True} def shared_memory_report() -> str: try: stat = os.statvfs("/dev/shm") return f"{stat.f_blocks * stat.f_frsize / 2**30:.1f} GiB" except OSError: return "unavailable" @contextlib.asynccontextmanager async def lifespan(_: FastAPI): global http log(f"/dev/shm: {shared_memory_report()}") await asyncio.to_thread(store.load_presets) if os.environ.get("DEMO_PREVIEW") == "1": engine.phase = "UI preview · GPU engine offline" else: import zerogpu_backend # app.py imports it first, so the weights are loaded http = zerogpu_backend.EngineClient() engine.use_in_process() transcriber.start() log("generation runs in this process, one ZeroGPU call per take") if os.environ.get("ENGINE_WARM_UP", "1") == "1" and hasattr(zerogpu_backend, "warm_up"): voice = next(iter(store.presets.values()), None) reference = {"data": voice.b64, "media_type": "audio/flac", "text": voice.transcript} if voice else None threading.Thread(target=zerogpu_backend.warm_up, args=(reference,), daemon=True).start() try: yield finally: if http is not None: await http.aclose() app = FastAPI(title="the engine", lifespan=lifespan) from starlette.middleware.sessions import SessionMiddleware app.add_middleware(SessionMiddleware, secret_key=secrets.token_hex(32)) # ─── Routes ─────────────────────────────────────────────────────────────────── @app.get("/favicon.ico", include_in_schema=False) async def favicon() -> Response: return Response(status_code=204) @app.get("/api/status") async def status(identity: Identity = Depends(identify)) -> dict: ttfas = sorted(recent_ttfa) return { "user": public_user(identity), "engine": engine.status(), "transcriber": transcriber.state, "gpu": await asyncio.to_thread(gpu_info), "load": {"active": gate.active, "waiting": len(gate.waiting), "slots": gate.slots}, "stats": { "takes": stats["takes"], "audio_s": round(stats["audio_s"]), "ttfa_ms_p50": round(ttfas[len(ttfas) // 2]) if ttfas else None, }, "limits": { "max_text_chars": MAX_TEXT_CHARS, "max_reference_s": REF_MAX_SECONDS, "max_audio_s": round(MAX_NEW_TOKENS * 2048 / SAMPLE_RATE), }, "voices": [voice.public() for voice in store.presets.values()], } @app.get("/api/voices/{voice_id}.wav", dependencies=[Depends(identify)]) async def voice_audio(voice_id: str) -> Response: voice = store.voice(voice_id) if voice is None: raise HTTPException(404, "That voice has expired. Add it again.") return Response(voice.wav, media_type="audio/wav", headers={"Cache-Control": "max-age=3600"}) @app.get("/api/takes/{take_id}.wav", dependencies=[Depends(identify)]) async def take_audio(take_id: str) -> Response: take = store.take(take_id) if take is None: raise HTTPException(404, "That take has expired.") return Response( wav_bytes(pcm_to_float(take.pcm), take.sample_rate), media_type="audio/wav", headers={"Content-Disposition": f'attachment; filename="engine-{take_id}.wav"'}, ) async def register_voice(wav: np.ndarray, sample_rate: int, label: str, fallback_text: str = "") -> dict: wav, cut = await asyncio.to_thread(tidy_reference, wav, sample_rate) duration = wav.size / SAMPLE_RATE if duration < REF_MIN_SECONDS: raise HTTPException(400, "That clip is too short. Record at least a second of speech.") transcript, language, transcribed = fallback_text, None, False if transcriber.state == "ready": try: transcript, language = await asyncio.to_thread(transcriber.transcribe, wav, SAMPLE_RATE) transcribed = True except Exception as error: log(f"transcription failed: {error!r}") audio = wav_bytes(wav, SAMPLE_RATE) voice = Voice( id=hashlib.sha1(audio).hexdigest()[:16], label=label, wav=audio, transcript=transcript, duration=duration, language=language, ) store.add_voice(voice) return {**voice.public(), "trimmed": cut, "transcribed": transcribed} @app.post("/api/voices", dependencies=[Depends(identify)]) async def add_voice(audio: UploadFile = File(...)) -> dict: data = await audio.read(UPLOAD_MAX_BYTES + 1) if len(data) > UPLOAD_MAX_BYTES: raise HTTPException(400, "That file is too big. Use a clip under 25 MB.") try: wav, sample_rate = await asyncio.to_thread(decode_audio, data) except Exception: raise HTTPException(400, "Couldn't read that audio. Try a WAV or MP3 file.") if not wav.size or not np.isfinite(wav).all() or float(np.abs(wav).max()) == 0.0: raise HTTPException(400, "That recording is silent. Check the microphone and try again.") label = Path(audio.filename or "").stem[:24] or "My voice" return await register_voice(wav, sample_rate, label) class FromTake(BaseModel): take_id: str label: str | None = Field(default=None, max_length=40) @app.post("/api/voices/from-take", dependencies=[Depends(identify)]) async def voice_from_take(body: FromTake) -> dict: take = store.take(body.take_id) if take is None: raise HTTPException(404, "That take has expired. Generate it again.") label = body.label or (take.choice.character.name if take.choice.character else "Found voice") return await register_voice( pcm_to_float(take.pcm), take.sample_rate, label, fallback_text=spoken_text(take.line) ) # ─── Streaming speech: text in, audio out, context token ────────────────────── class SpeechTurn(BaseModel): text: str codes: str | None = None # base64 int16 [frames, num_codebooks] audio: str | None = None # base64 clip, for backends that can only return audio class SpeechBody(BaseModel): text: str voice: str | None = None language: str | None = None history: list[SpeechTurn] = Field(default_factory=list) max_frames: int = Field(default=MAX_NEW_TOKENS_CAP, ge=1, le=MAX_NEW_TOKENS_CAP) temperature: float = Field(default=1.0, ge=0.1, le=1.5) top_p: float = Field(default=0.9, ge=0.1, le=1.0) top_k: int = Field(default=0, ge=0, le=ALL_CANDIDATES) def decode_codes(value: str, num_codebooks: int) -> list[list[int]]: """base64 int16 [frames, codebooks] → the [codebooks, frames] grid the engine takes.""" raw = base64.b64decode(value) frames = len(raw) // 2 // num_codebooks if frames <= 0: raise HTTPException(400, "history codes are empty") grid = np.frombuffer(raw, dtype=" tuple[dict, "Voice", str]: """The engine request for one chunk, shared by /speech and /speech/start.""" voice = store.voice(body.voice or "clone1") if voice is None: raise HTTPException(404, "unknown voice") text = clean_text(body.text, "script", MAX_TEXT_CHARS) references: list[dict] = [{"data": voice.b64, "media_type": "audio/flac", "text": voice.transcript}] history: list[str] = [] for turn in body.history[-HISTORY_TAKES:]: if not turn.codes and not turn.audio: raise HTTPException(400, "a history turn needs codes or audio") clean_text(turn.text, "history text", MAX_TEXT_CHARS) if turn.codes: references.append({"vq_codes": decode_codes(turn.codes, 10), "text": turn.text}) else: references.append({"data": turn.audio, "media_type": "audio/wav", "text": turn.text}) history.append(turn.text) envelope = {"text": text, "voice_references": 1, "history": history} payload = { "input": ENVELOPE_MARKER + json.dumps(envelope, ensure_ascii=False), "references": references, "stream": True, "response_format": "pcm", "stream_format": "sse", "temperature": body.temperature, "top_p": body.top_p, "top_k": body.top_k or ALL_CANDIDATES, "max_new_tokens": body.max_frames, } return payload, voice, text @app.post("/speech") async def speech(body: SpeechBody, request: Request) -> StreamingResponse: """One chunk of text in, acting audio out, plus the codes that continue it.""" from gradio import Request as GradioRequest from gradio.context import LocalContext if LocalContext.request.get(None) is None: LocalContext.request.set(GradioRequest(request=request)) if engine.state != "ready": raise HTTPException(503, "the engine is still warming up") payload, voice, text = speech_payload(body) async def stream(): started = time.perf_counter() yield sse({"type": "start", "sample_rate": SAMPLE_RATE, "voice": voice.id}) yield sse({"type": "segment", "index": 0, "text": text}) audio_bytes = 0 try: response = await http.send( http.build_request("POST", f"{ENGINE_URL}/v1/audio/speech", json=payload), stream=True ) if response.status_code != 200: message = engine_error_message(await response.aread()) yield sse({"type": "error", "message": message}) return sample_rate = int(response.headers.get("x-sample-rate", SAMPLE_RATE)) progress = {"audio_at": time.perf_counter(), "has_audio": False} async with contextlib.aclosing(engine_lines(response, started, progress)) as lines: async for raw in lines: if raw is None: yield sse({"type": "heartbeat"}) continue if not raw.startswith("data: "): continue event = json.loads(raw[6:]) kind = event.get("type") if kind == "speech.audio.delta": progress.update(audio_at=time.perf_counter(), has_audio=True) chunk = event["audio"] audio_bytes += len(base64.b64decode(chunk)) if isinstance(chunk, str) else len(chunk) yield sse({"type": "audio", "pcm": chunk}) elif kind == "speech.audio.done": yield sse({ "type": "done", "voice": voice.id, "sample_rate": sample_rate, "frames": event.get("frames"), "finish_reason": event.get("finish_reason"), "audio_s": round(audio_bytes / 2 / sample_rate, 2), "codes": event.get("codes"), "timings_ms": event.get("timings"), }) return elif kind == "error": yield sse({"type": "error", "message": str(event.get("message"))[:300]}) return except asyncio.TimeoutError: yield sse({"type": "error", "message": "the engine timed out"}) except Exception as error: # noqa: BLE001 log(f"speech failed: {error!r}") yield sse({"type": "error", "message": "the engine failed"}) return StreamingResponse( stream(), media_type="text/event-stream", headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"}, ) # ─── Start / stream split ───────────────────────────────────────────────────── # # The create call is the one that acquires the GPU, so it has to arrive from a # fresh residential IP: that is what ZeroGPU charges. The stream carries no GPU # work, so it can be read straight from the Space. Splitting them lets the proxy # spend one ScrapeOps session per take and pull the audio over the open internet. TAKE_TTL_S = float(os.environ.get("SPEECH_TAKE_TTL_S", 300)) MAX_TAKES = int(os.environ.get("SPEECH_MAX_TAKES", 8)) _takes: "collections.OrderedDict[str, TakeBuffer]" = collections.OrderedDict() class TakeBuffer: """Events of one take, produced by a thread, drained by a GET.""" def __init__(self) -> None: self._events = collections.deque() self._cond = threading.Condition() self._closed = False self.created = time.time() def push(self, event: dict) -> None: with self._cond: self._events.append(event) self._cond.notify_all() def close(self) -> None: with self._cond: self._closed = True self._cond.notify_all() def get(self, timeout: float): """Next event, a heartbeat when idle, or None once finished and drained.""" deadline = time.time() + timeout with self._cond: while not self._events and not self._closed: remaining = deadline - time.time() if remaining <= 0: return {"type": "heartbeat"} self._cond.wait(remaining) if self._events: return self._events.popleft() return None @property def expired(self) -> bool: return time.time() - self.created > TAKE_TTL_S def _prune_takes() -> None: for take_id in [key for key, value in _takes.items() if value.expired]: _takes.pop(take_id, None) while len(_takes) > MAX_TAKES: _takes.popitem(last=False) @app.post("/speech/start") async def speech_start(body: SpeechBody, request: Request) -> dict: """Begin one take; the caller streams it from /speech/take/{id}.""" import contextvars from gradio import Request as GradioRequest from gradio.context import LocalContext if LocalContext.request.get(None) is None: LocalContext.request.set(GradioRequest(request=request)) if engine.state != "ready": raise HTTPException(503, "the engine is still warming up") payload, voice, text = speech_payload(body) take_id = secrets.token_urlsafe(12) buffer = TakeBuffer() _prune_takes() _takes[take_id] = buffer context = contextvars.copy_context() def produce() -> None: import zerogpu_backend budget = zerogpu_backend.gpu_seconds(payload) - 3 try: for event in zerogpu_backend.gpu_speech(payload, budget): buffer.push(event) except BaseException as error: # noqa: BLE001 buffer.push({"type": "error", "message": repr(error)[:300]}) finally: buffer.close() threading.Thread(target=context.run, args=(produce,), daemon=True).start() log(f"take {take_id} started on {voice.id} ({len(text)} chars)") return {"take_id": take_id, "voice": voice.id, "chars": len(text)} @app.get("/speech/take/{take_id}") async def speech_take(take_id: str) -> StreamingResponse: buffer = _takes.get(take_id) if buffer is None: raise HTTPException(404, "unknown or expired take") async def stream(): audio_bytes = 0 while True: event = await asyncio.to_thread(buffer.get, 15.0) if event is None: return kind = event.get("type") if kind == "speech.audio.delta": chunk = event["pcm"] audio_bytes += len(chunk) yield sse({"type": "audio", "pcm": base64.b64encode(chunk).decode("ascii")}) elif kind == "speech.audio.done": yield sse({"type": "done", "frames": event.get("frames"), "finish_reason": event.get("finish_reason"), "audio_s": round(audio_bytes / 2 / SAMPLE_RATE, 2), "codes": event.get("codes"), "timings_ms": event.get("timings")}) return elif kind == "error": yield sse({"type": "error", "message": str(event.get("message"))[:300]}) return else: yield sse({"type": "heartbeat"}) return StreamingResponse(stream(), media_type="text/event-stream", headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"}) class VoiceBody(BaseModel): mode: Literal["clone", "design", "random"] id: str | None = None transcript: str | None = None name: str | None = None description: str | None = None scene: str | None = None class GenerateBody(BaseModel): text: str voice: VoiceBody retake_of: str | None = None note: str | None = None temperature: float = Field(default=1.0, ge=0.1, le=1.5) top_p: float = Field(default=0.9, ge=0.1, le=1.0) top_k: int = Field(default=0, ge=0, le=ALL_CANDIDATES) # 0: no cap def clean_text(value: str | None, what: str, limit: int, *, single_line: bool = False) -> str: value = (value or "").replace("\r\n", "\n").strip() if single_line: value = re.sub(r"\s+", " ", value) if not value: raise HTTPException(400, f"Add the {what} first.") if len(value) > limit: raise HTTPException(400, f"Keep the {what} under {limit} characters.") if STRUCTURAL.search(value) or ENVELOPE_MARKER in value: raise HTTPException(400, f"Remove the <|…|> tokens from the {what}; this page writes those itself.") return value def resolve_voice(body: VoiceBody) -> VoiceChoice: if body.mode == "clone": voice = store.voice(body.id or "") if voice is None: raise HTTPException(404, "That voice has expired. Upload or record it again.") transcript = clean_text( body.transcript if body.transcript is not None else voice.transcript, "reference transcript", MAX_TRANSCRIPT_CHARS, single_line=True, ) return VoiceChoice(mode="clone", voice=voice, transcript=transcript) if body.mode == "design": name = clean_text(body.name or "Narrator", "character name", MAX_NAME_CHARS, single_line=True) if ":" in name: raise HTTPException(400, "Leave ':' out of the character name.") description = clean_text(body.description, "voice description", MAX_DESCRIPTION_CHARS, single_line=True) scene = (body.scene or "").strip() scene = clean_text(scene, "scene", MAX_SCENE_CHARS, single_line=True) if scene else None return VoiceChoice(mode="design", character=Character(name, description, scene)) return VoiceChoice(mode="random") def retake_text(line: str, note: str) -> str: """Give the new performance its own direction; history holds the old one. Repeating the original script's acting tags makes them compete with the director's new note. Preserve the spoken words, replace the acting cues, and let the user decide whether their note is absolute or comparative. """ note = clean_text(re.sub(r"[\[\]<>|]", "", note), "note", MAX_NOTE_CHARS, single_line=True) words = clean_text(spoken_text(line), "spoken script", MAX_TEXT_CHARS) # A single short prefix is often drowned out by the reference performance. # Keep the note active through each sentence rather than only at the opening. sentences = re.split(r"(?<=[.!?。!?])\s+", words) directed = " ".join(f"[{note}] {sentence}" for sentence in sentences) return f"Direction: {note}\nLine: {directed}" def engine_request(choice: VoiceChoice, turn_text: str, history: list[Take], body: GenerateBody) -> dict: references = [] if choice.mode == "clone": references.append({"data": choice.voice.b64, "media_type": "audio/flac", "text": choice.transcript}) elif history: voice = choice.directing_voice if voice is None: raise HTTPException(500, "The generated voice reference is missing for this retake.") references.append({"data": voice.b64, "media_type": "audio/flac", "text": voice.transcript}) for take in history: references.append({ "data": base64.b64encode(flac_bytes(np.frombuffer(take.pcm, dtype=" str: try: payload = json.loads(raw) message = payload.get("error", {}).get("message") or payload.get("message") or payload.get("detail") except (ValueError, AttributeError): message = None message = str(message or raw or "The engine refused this request.") return message.removeprefix("the engine request:").strip()[:400] def sse(payload: dict) -> str: return f"data: {json.dumps(payload)}\n\n" @app.post("/api/generate") async def generate( body: GenerateBody, identity: Identity = Depends(identify), request: Request = None ) -> StreamingResponse: if request is not None: # API callers: let ZeroGPU see this request's headers (Gradio sets this # itself for the Studio and Workflow). from gradio import Request as GradioRequest from gradio.context import LocalContext if LocalContext.request.get(None) is None: LocalContext.request.set(GradioRequest(request=request)) if engine.state != "ready": raise HTTPException(503, "the engine is still warming up. It will be ready in a moment.") line = clean_text(body.text, "script", MAX_TEXT_CHARS) history: list[Take] = [] if body.retake_of: previous = store.take(body.retake_of) if previous is None: raise HTTPException(404, "That take has expired. Generate a fresh one.") choice = previous.choice if choice.mode != "clone" and choice.directing_voice is None: choice = replace(choice, directing_voice=Voice( id=f"take-{previous.id}", label=choice.character.name if choice.character else "Surprise", wav=wav_bytes(pcm_to_float(previous.pcm), previous.sample_rate), transcript=spoken_text(previous.line) or previous.line, duration=previous.seconds, )) chain = [take for take in map(store.take, previous.history) if take is not None] history = chain[max(0, len(chain) - (HISTORY_TAKES - 1)):] + [previous] note = clean_text(body.note, "note", MAX_NOTE_CHARS, single_line=True) turn_text = retake_text(line, note) else: choice = resolve_voice(body.voice) turn_text = line if len(gate.waiting) >= MAX_QUEUE: raise HTTPException(503, "The stage is packed right now. Try again in a minute.") refusal = limits.check(identity.key) if refusal: raise HTTPException(429, refusal) payload = engine_request(choice, turn_text, history, body) take_id = secrets.token_urlsafe(9) async def stream(): slot = limits.admit(identity.key) if isinstance(slot, str): yield sse({"type": "error", "message": slot}) return ticket = gate.enqueue() queued_at = time.perf_counter() try: while not ticket.granted.is_set(): yield sse({"type": "queued", "position": gate.position(ticket)}) try: await asyncio.wait_for(ticket.granted.wait(), timeout=1.0) except asyncio.TimeoutError: pass admitted = time.perf_counter() queue_ms = (admitted - queued_at) * 1000 started = time.perf_counter() timings = {"queue": round(queue_ms)} pcm = bytearray() ttfa_ms = None finish = None sent_audio = False async def audio_events(kept, *, flush=False): nonlocal ttfa_ms, sent_audio buffered.extend(kept) if not buffered: return if not sent_audio and not flush and len(buffered) < START_BUFFER_S * sample_rate * 2: return kept = bytes(buffered) buffered.clear() if not sent_audio: ttfa_ms = (time.perf_counter() - started) * 1000 timings["opening_validated"] = round(ttfa_ms) yield sse({"type": "start", "take_id": take_id, "sample_rate": sample_rate, "queue_ms": round(queue_ms)}) pcm.extend(kept) packet_bytes = max(2, round(AUDIO_PACKET_S * sample_rate) * 2) for offset in range(0, len(kept), packet_bytes): packet = kept[offset:offset + packet_bytes] chunk = {"type": "audio", "pcm": base64.b64encode(packet).decode("ascii")} if not sent_audio: chunk["ttfa_ms"] = round(ttfa_ms) timings["first_packet_after_admission"] = round((time.perf_counter() - admitted) * 1000) sent_audio = True yield sse(chunk) for attempt in range(2): buffered = bytearray() guard = None response = None attempt_started = time.perf_counter() opening = asyncio.ensure_future(http.send( http.build_request("POST", f"{ENGINE_URL}/v1/audio/speech", json=payload), stream=True )) try: while not opening.done(): remaining = min(ENGINE_START_TIMEOUT_S - (time.perf_counter() - attempt_started), TAKE_TIMEOUT_S - (time.perf_counter() - started)) if remaining <= 0: raise asyncio.TimeoutError if not (await asyncio.wait({opening}, timeout=min(KEEPALIVE_S, remaining)))[0]: yield ": keepalive\n\n" response = opening.result() if response.status_code != 200: message = engine_error_message(await asyncio.wait_for(response.aread(), ENGINE_IDLE_TIMEOUT_S)) log(f"engine refused {take_id}: {message}") yield sse({"type": "error", "message": message}) return sample_rate = int(response.headers.get("x-sample-rate", SAMPLE_RATE)) guard = SilenceGuard(sample_rate) progress = {"audio_at": attempt_started, "has_audio": False} async with contextlib.aclosing(engine_lines(response, started, progress)) as lines: async for raw in lines: if raw is None: yield ": keepalive\n\n" continue if not raw.startswith("data: "): continue event = json.loads(raw[6:]) kind = event.get("type") if kind == "speech.audio.delta": timings.setdefault("engine_first_audio", round((time.perf_counter() - started) * 1000)) progress.update(audio_at=time.perf_counter(), has_audio=True) kept = guard.feed(base64.b64decode(event["audio"])) async for chunk in audio_events(kept): yield chunk if guard.stalled: finish = "stall" break elif kind == "speech.audio.done": finish = event.get("finish_reason") or "stop" if event.get("timings"): timings["engine"] = event["timings"] break # A done event is final; do not wait for socket EOF. elif kind == "error": message = engine_error_message(json.dumps(event)) log(f"engine failed {take_id}: {message}") yield sse({"type": "error", "message": message}) return except (asyncio.TimeoutError, httpx.TimeoutException): finish = "timeout" finally: # Cancellation can race with response headers arriving. if not opening.done(): opening.cancel() await asyncio.gather(opening, return_exceptions=True) if response is None and not opening.cancelled() and opening.exception() is None: response = opening.result() if response is not None: await response.aclose() if guard is not None: tail = guard.feed(b"", final=True) # A failed opening's tail must not push it over the playback # threshold just before it is discarded and retried. if sent_audio or finish in {"stop", "timeout"}: async for chunk in audio_events(tail): yield chunk # Normal short lines finish without filling the startup buffer. # A stalled fragment is discarded only if nothing has gone on air. if finish in {"stop", "timeout"} or sent_audio: async for chunk in audio_events(b"", flush=True): yield chunk if pcm: break if attempt == 0 and finish != "timeout": log(f"take {take_id}: silent or incomplete opening, retrying once") yield sse({"type": "retrying", "message": "The actor did not finish the opening. Trying once more…"}) finish = None else: message = ("The actor did not respond in time. Try again." if finish == "timeout" else "No complete opening came out after two attempts. Try a different voice or direction.") yield sse({"type": "error", "message": message}) return warning = { "stall": "The actor went quiet before ending the take. Part of the line may be missing; try a retake.", "length": "The actor reached the time limit. The line may be incomplete; try a retake or a shorter script.", "timeout": "The actor stopped responding. This take may be incomplete; try a retake.", }.get(finish) if finish is None: finish = "incomplete" warning = "The stream ended before the actor finished. This take may be incomplete; try a retake." total_ms = (time.perf_counter() - started) * 1000 seconds = len(pcm) / 2 / sample_rate trimmed_s = guard.dropped / sample_rate store.add_take(Take( id=take_id, choice=choice, line=line, turn_text=turn_text, history=[take.id for take in history], pcm=bytes(pcm), sample_rate=sample_rate, )) stats["takes"] += 1 stats["audio_s"] += seconds recent_ttfa.append(ttfa_ms or total_ms) log(f"take {take_id}: {choice.mode}, {len(history)} earlier, {seconds:.1f}s audio, " f"{trimmed_s:.1f}s silence trimmed, ttfa {ttfa_ms or 0:.0f}ms, " f"total {total_ms:.0f}ms, queue {queue_ms:.0f}ms, {finish}") yield sse({ "type": "done", "take_id": take_id, "ttfa_ms": round(ttfa_ms or total_ms), "total_ms": round(total_ms), "queue_ms": round(queue_ms), "timings_ms": timings, "audio_s": round(seconds, 2), "speed": round(seconds / (total_ms / 1000), 2) if total_ms > 0 else None, "finish_reason": finish, "warning": warning, "retry_count": attempt, "trimmed_s": round(trimmed_s, 1), "turn_text": turn_text, }) except httpx.HTTPError as error: log(f"engine connection failed for {take_id}: {error!r}") yield sse({"type": "error", "message": "Lost the connection to the engine. Try again."}) finally: gate.leave(ticket) limits.release(identity.key, slot) return StreamingResponse( stream(), media_type="text/event-stream", headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"}, ) from ui import mount_apps app = mount_apps(sys.modules[__name__]) app.add_middleware(StreamingGZipMiddleware) if __name__ == "__main__": uvicorn.run(app, host="0.0.0.0", port=PORT, log_level="info", access_log=False)