j$dZddlmZddlZddlZddlZddlmZddlm Z m Z eejjddZ d d d d d iZGddZGddZddZeddZy)uWhere the spoken words come from — one interface, two engines. WHY THIS EXISTS. Commentary audio was synthesised by exactly one vendor, reached by exactly one code path, and the coupling cost more than it looked. Three separate problems, one shape: * **Money.** Google's tier auto-selection lands on the most expensive voice that exists, and on 13 August 2026 — one championship day — Text-to-Speech billed **kr 7,480 in Studio voices alone**, against an expected kr 277 for the whole estate. * **Availability.** `TextToSpeechClient()` needs Application Default Credentials at *construction*, so a box without them lost the whole synthesis thread rather than degrading. The office deployment has no ADC and is not getting any. * **Sovereignty.** Every spoken line left the building to be spoken. That is a vendor dependency in the one part of the broadcast that is pure computation. A local engine answers all three at once, so the choice belongs behind an interface rather than an `if` in the middle of the cache loop. WHAT AN ENGINE OWES. Two things and no more: which voice each commentary role speaks in, and how a line becomes audio bytes. Everything else — the queue, the circuit breaker, the digest key, the index — stays in `live_cache.py`, which is engine-agnostic and should remain so. THE ONE-VOICE RULE APPLIES TO BOTH. The broadcast standard is exactly **one male and one female voice per language** — Male Caller for facts, Female Analyst for the story — held stable across a whole programme, because a voice that drifts between renders reads as a different person arriving mid-race. Both engines pin their set deliberately and both let the same `TTS_VOICE__` environment variables override it, so an operator learns one convention rather than two. ) annotationsN)Path) LANG_CODE pick_voicesVR_PIPER_VOICESz/srv/vr/voicesenzen_GB-alan-mediumzen_GB-jenny_dioco-mediumz"en_GB-northern_english_male-medium)leadtacticalrulesc<eZdZdZdZdZddZd dZ d dZy) GoogleEngineuUCloud synthesis, unchanged — the SSML path and its plain-text retry both live here.z.mp3googlecHddlm}||_|j|_y)Nr) texttospeech) google.cloudr_ttsTextToSpeechClient_client)selfrs >/srv/vr/releases/25f455cc4e75/v3/broadcast/live/tts/engines.py__init__zGoogleEngine.__init__<s-  #668 c.t|j|S)N)rr)rlangs rvoiceszGoogleEngine.voicesBs4<<..rc|j}|jt||}|j|jj |} |r|j |n|j |} |jj| || jS#t$r>|s|jj|j ||| jcYSwxYw)N) language_codename)audio_encoding speaking_rate)ssml)text)inputvoice audio_config) rVoiceSelectionParamsr AudioConfig AudioEncodingMP3SynthesisInputrsynthesize_speech audio_content Exception) rrpayloadis_ssmlspoken voice_nameratettsr%configinps r synthesizezGoogleEngine.synthesizeEsii((yZ(X0A0A0E0EUYZ29c  g .s?Q?QW^?Q?_ g<<115W]1^ll l g<<11((f(5UQW2YYfYf g  gs5'BAC$#C$N)returnNonerstrr8zdict[str, str]rr;r/r;r0boolr1r;r2r;r3floatr8bytes) __name__ __module__ __qualname____doc__suffixrrrr7rrrr6s:_ F D9 /g"g*/g49grrcTeZdZdZdZdZdedd dZd dZd dZ dd Z dd Z y) PiperEngineuLocal neural synthesis on the CPU, no network and no credential. Measured on the producer VM (2 vCPU) with real commentary lines: **~16x realtime**, so a race day's three to four hours of speech is about thirteen minutes of compute. The same models ran at 33.7x on the render host's 24 threads — quote whichever machine will actually do the work. WHAT IS GIVEN UP, stated rather than glossed: Piper takes **plain text**, so the SSML pronunciation layer does not reach it. That layer carries per-language fixes — Swedish `banan` (the course, not the fruit) and the hard `g` in `gippa` — and for English exactly one entry, the broadcast sense of *live* as /laɪv/. So the English loss today is one word, and the Swedish dictionary matters only when Swedish output is switched back on. This is the same degradation the cloud path already performs whenever SSML is rejected, not a new failure mode. z.wavpiperN) voice_diron_logc|xst|_||_i|_|jj st d|jdy)Nzno Piper voice directory at z (set $VR_PIPER_VOICES)) VOICE_DIR_dir_log_loadedis_dir RuntimeError)rrIrJs rrzPiperEngine.__init__hsK*  *, yy!!=dii[H_`a a"rc&|j|dz S)Nz.onnx)rM)rr2s r_modelzPiperEngine._modelosyyj\///rc tj|}|std||j}i}|j D]\}}t j jd|d|j}|r%|j|jr|||<`|r$|jdt||d|d||||<|jDcgc]$}|j|jr#|&} }| rddlm}|j t |j |}||j|<|S)Nr) PiperVoice)rOr\rHroloadr;rS)rr2r%ros r_voicezPiperEngine._voicesO   , = (OOC J(?$@AE',DLL $ rc0ddlm}||rd|z nd}tj} t j | d5} |j |j|| |ddd| jS#1swY| jSxYw)Nr)SynthesisConfigg?) length_scalewb) syn_config) rHrsioBytesIOwaveopenrqsynthesize_wavgetvalue) rrr/r0r1r2r3rsr5bufferouts rr7zPiperEngine.synthesizes) !DcDjcJ YYvt $ S KK # 2 2636 2 R S   S  s $A==B)rIz Path | Noner8r9)r2r;r8rr:)r2r;r<) r@rArBrCrDrprintrrSrrqr7rErrrGrGVsK F D37b0^8  !" !*/ !49 !rrGc.d|d|jS)NrUrV)r]) lang_upperrjs rrbrbs  |1TZZ\N 33rrJctjjdxsdjj }|dk(r t |S|dk(r t Std|d)uThe deployment's engine, from `$VR_TTS_ENGINE`. Defaults to **google**, so no existing deployment changes behaviour by upgrading. Set `VR_TTS_ENGINE=piper` on a box that has the models and no credential — which is the office deployment exactly. VR_TTS_ENGINErrHrzunknown $VR_TTS_ENGINE z (expected 'google' or 'piper'))r_r`r\striplowerrGrrQ)rJchoices r make_enginersfjjnn_-9 @ @ B H H JF &)) ~ 0 :YZ [[r)rr;rjr;r8r;)rC __future__rrwr_rypathlibrsynthrrr`r\rLr[rrGrbrrrErrrs6# )  13CD E  #.5 gg@N!N!b4 \r