Calliope voice auditions

2026-10-20 · every Google voice tier, what each can do, and which settings matter. Ratings and notes save in this browser; tap Copy results when done and paste them back to me.

Every tier, same line

Two voices per tier, all saying: “Okay, I looked into it. The short answer is yes, but there's a catch: the cheaper plan drops the offline mode, and honestly, that's the part you'd miss. Want me to compare the two?” “Ready in” is the time for the whole clip from our server (a call starts speaking sooner, sentence by sentence). ≈/mo is the estimate for your last 30 days of speech; * = worst case, the free 4M pool is shared with the podcasts.

Standard $4 / 1M chars ≈$0.52*/mo
Oldest, most robotic. Full SSML.
Standard-H
ready in 0.93 s
Standard-J
ready in 0.96 s
WaveNet $4 / 1M chars ≈$0.52*/mo
What Calliope used first. Full SSML.
Wavenet-D
ready in 0.79 s
Wavenet-F
ready in 0.87 s
Neural2 $16 / 1M chars ≈$0/mo
Smoother classic voice. Full SSML + word timings.
Neural2-F
ready in 0.85 s
Neural2-J
ready in 0.54 s
Studio $160 / 1M chars ≈$0/mo
Narration voices. Most SSML. Slower to start.
Studio-O
ready in 1.62 s
Studio-Q
ready in 1.59 s
Chirp HD $30 / 1M chars ≈$0/mo
Conversational. No SSML.
Chirp-HD-F
ready in 2.69 s
Chirp-HD-D
ready in 2.77 s
Chirp 3 HD $30 / 1M chars ≈$0/mo
30 voices. Pause markup, custom pronunciations, SSML.
Chirp3-HD-Sulafat
ready in 1.66 s
Chirp3-HD-Achird
ready in 1.6 s
Gemini 3.8 Flash $16 (→$32 in 2027) / 1M chars ≈$2.10/mo no speed setting
Newest. Style direction, <tags>, voice design. No speed setting.
Kore
ready in 4.33 s
Puck
ready in 4.14 s
Gemini 3.8 Flash-Lite $11 (→$22 in 2027) / 1M chars ≈$1.44/mo no speed setting
Cheaper 3.8. Same controls.
Leda
ready in 3.63 s
Orus
ready in 3.41 s
Gemini 2.5 Flash $18 / 1M chars ≈$2.36/mo
Style prompt + [tags].
Kore
ready in 9.14 s
Puck
ready in 6.63 s
Gemini 2.5 Flash-Lite $18 / 1M chars ≈$2.36/mo
Style prompt + [tags].
Leda
ready in 6.25 s
Orus
ready in 7.25 s
Gemini 2.5 Pro $36 / 1M chars ≈$4.72/mo
Slowest to start (~4.5 s).
Aoede
ready in 11.02 s
Charon
ready in 13.89 s
Gemini 3.1 Flash $36 / 1M chars ≈$4.72/mo
Style prompt + [tags].
Zephyr
ready in 7.19 s
Fenrir
ready in 9.25 s
Gemini 3.8 — direction and tags

These were checked only by length and status code, not by ear. Tell me whether each one actually did what it says.

Plain sentence, no direction.
3.8 Flash-Lite Reference for the next five.
style: “whispering softly”
3.8 Flash-Lite Is it actually a whisper?
style: “whispering softly”
3.8 Flash Same on the bigger model.
style: “thrilled and excited”
3.8 Flash-Lite Audible excitement, not just faster?
style: “speaking rapidly”
3.8 Flash-Lite Usable as a speed control? (2.4 s vs 3.4 s)
style: “speaking slowly”
3.8 Flash-Lite Natural, or dragged? (6.0 s)
“Well <sigh> that took a while. <short pause> But it worked <laugh> finally.”
3.8 Flash-Lite Do you hear a sigh, a pause and a laugh — and NOT the words “sigh”/“laugh”?
Same line with <tags>.
3.8 Flash Same check on Flash.
Same line with the OLD [square] tags.
3.8 Flash-Lite Does it say “sigh” out loud? (Docs say 3.8 wants <angle> tags.)
One word in CAPITALS.
3.8 Flash-Lite Is that word stressed?
Two speakers, one interjecting “|mm-hm|” mid-sentence.
3.8 Flash-Lite Natural overlap? (two-voice podcasts/dialogue)
A voice DESIGNED from a text description (created, used, deleted).
3.8 Flash-Lite Could this be Calliope's own voice?
A voice from the 2,000-voice Extended Library (“puck”).
3.8 Flash-Lite Different from the 30 presets?
Gemini 2.5 / 3.1 — style prompt and [tags]

These were checked only by length and status code, not by ear. Tell me whether each one actually did what it says.

Plain sentence.
2.5 Flash Reference.
prompt: “Whisper this very quietly”
2.5 Flash Whisper?
Inline [whispering] tag.
2.5 Flash Whisper, or does it say “whispering”?
speakingRate 1.5 (native).
2.5 Flash Clean at 1.5×?
prompt “amused” + [laughing].
3.1 Flash A real laugh?
[short pause] / [long pause] tags.
3.1 Flash Pauses, not words?
[scared] — docs say this kind of tag is SPOKEN.
3.1 Flash Does it say “scared”?
Chirp 3 HD — pace, pauses, pronunciation, SSML

These were checked only by length and status code, not by ear. Tell me whether each one actually did what it says.

speakingRate 1.0
Chirp 3 HD Reference.
speakingRate 1.5
Chirp 3 HD Clean at 1.5×?
Sentence, no pause.
Chirp 3 HD Reference.
[pause long] sent as markup.
Chirp 3 HD A real pause?
[pause long] sent as plain text (what the app does today).
Chirp 3 HD Does it read “pause long” aloud?
“Calliope”, default pronunciation.
Chirp 3 HD Reference.
“Calliope” with an IPA pronunciation (kə-LY-ə-pee).
Chirp 3 HD Pronounced right?
SSML <break time="2s"/>
Chirp 3 HD A 2-second pause?
Classic tiers — SSML and effects

These were checked only by length and status code, not by ear. Tell me whether each one actually did what it says.

SSML break + <emphasis> + <prosody>.
WaveNet Each effect audible?
effectsProfileId: telephony-class-application
WaveNet Sounds like a phone line?
SSML break + prosody rate.
Studio Pause and slower part?
SSML <mark>s; Google returned the time of each (0.37 s, 0.84 s).
Neural2 Nothing to hear — this is what would let text highlight as she speaks.
speakingRate 1.0
Chirp HD Reference.
speakingRate 1.5
Chirp HD Clean at 1.5×?
Speed we'd do ourselves

3.8 rejects a speed setting, so the shim would stretch the audio after Google returns it. These are the 3.8 Flash “Kore” clip above, stretched on our server. The page-wide speed slider at the bottom does the same thing in your browser (pitch kept) — the app could do it on playback instead.

1.0× (original, 11.4 s)
1.15× (ffmpeg atempo)
1.3× (atempo)
1.3× (rubberband — better quality stretcher)
1.5× (atempo)
Settings worth using — my take

Today the app sends Google four things: the text, the voice, the audio format and a speed. Everything below is unused so far. In rough order of payoff for Calliope:

1. Let her perform: style + tags on Gemini 3.8

  • Style (speechMetadata.style): a short phrase like “warm, relaxed, a little wry” sent with each reply. It costs about 3 text tokens, effectively free. Google warns that long or changing style text makes the voice drift, so I'd keep it short and fixed per voice.
  • Inline tags in the text: <laugh> <sigh> <short pause> <whispers> and about 40 more, plus CAPITALS for emphasis. The idea: when a 3.8 voice is picked, Calliope's prompt tells her she may use them, and she directs her own delivery.
  • The text catch: the chat transcript and copy/export must strip the tags, or you'd read “<sigh>” on screen. Also, 3.8 reads its text word for word, so any instruction placed in the text gets spoken. Direction belongs in style, never in the text.
  • 3.8 wants <angle> tags; the older Gemini models use [square] ones. The shim should translate, so Calliope only learns one form. The feature clips above test whether a square tag gets read out on 3.8.

2. Pace: ask Gemini for it, don't stretch it — it's cheaper

Yes, asking 3.8 to speak faster lowers the bill. Google bills exactly 25 audio tokens per second of speech; we checked the billed count on every call:

3.8 Flash-Lite, same 2 sentencesAudioBilled audio tokensCost per reply
No style6.8 s169–171$0.00103
“speaking rapidly”4.5–4.9 s113–123$0.00069–0.00075 (−28 to −33%)
“speaking quickly but clearly”4.7–5.2 s118–129$0.00072–0.00079 (−24 to −30%)
  • At your usage that's about $1.44 → ~$1.00 a month on Flash-Lite, or $2.10 → ~$1.50 on Flash. It's real but small money; the bigger win is simply that she gets to the point faster.
  • Stretching the audio ourselves afterwards saves nothing: Google bills what it generated. So a “Pace” setting should be style text on 3.8 (cheaper, but coarse and varies run to run). On the other tiers it should be the native speed number. Our own stretch is only for an exact speed on 3.8. Playback speed in the app (the slider below) is the same as our own stretch, at zero server cost.
  • For the older Gemini models via Cloud TTS, the native speed setting did shorten the audio (4.0 s → 2.6 s at 1.5×). Google doesn't say whether it's billed before or after the speed-up, and those calls report no usage. The billing export will tell us after a day of use; I won't assume.

3. Faster first word: streaming

Today each sentence is synthesized in full before it plays. Streaming starts playing while Google is still generating. It's available for 3.8 (streamGenerateContent, raw audio chunks) and for Chirp 3 HD / Gemini 2.5–3.1 (Cloud TTS streaming, gRPC only). This is the biggest latency lever on the list, and the biggest build too: a new code path in the shim. Note: no SSML in streaming mode.

4. A voice that's hers: 3.8 voice design

Describe a voice in a paragraph; Google returns a reusable voice ID. A test voice was created, used and deleted fine (clip above). It replaces the 30 shared presets with one persona that stays consistent across replies. IDs are kept for a year after last use. There's also a 2,000-voice library with accent and persona tags, if we want a browsable picker beyond the 30.

5. Saying things right

  • Custom pronunciation (Chirp 3 HD): pin “Calliope”, names and acronyms to a phonetic spelling; it applies every time, no markup needed. On Gemini there's no equivalent; respelling in the text (“Kuh-LYE-uh-pee”) is the fallback.
  • Pause markup (Chirp 3 HD): [pause long] only works in the separate markup field. Sent as plain text, as we do today, it seems to be read out (clip above).
  • SSML (classic tiers, Studio, Chirp 3 HD): <say-as> for dates, numbers and phone numbers; <sub> to say “Doctor” for “Dr.”. Useful on the cheap tiers; moot on Gemini, which already reads naturally.

6. Phone-line and loudness polish

Volume gain and device profiles (handset, headphones, telephony EQ) are cheap one-field wins on the classic tiers. Unverified on Chirp 3 and Gemini. 8 kHz μ-law/A-law output exists on every API if Calliope ever answers real phone calls; 3.8 can return it directly, skipping our conversion.

7. Only on the old voices: word timings

<mark> + time-pointing returns the exact moment each marked word is spoken, which is what you'd need to highlight text in sync or drive the orb from words. It works only on Neural2/WaveNet; Chirp 3 returns nothing and Gemini has no such feature. Not worth choosing a voice for.

Safe to ignore

Gender (ignored once a voice is named) · long-audio synthesis (async to a storage bucket, not real-time) · safety-filter settings (off by default; one is deprecated) · language code on 3.8 (auto-detected) · temperature/top-p on 3.8 (rejected) · voice cloning (needs a consent recording) · custom-voice fields (enterprise-only).

Traps if we build any of this

  • Pitch: Chirp 3 HD returns an error for it; WaveNet, Neural2 and Gemini 2.5 accept it. It would have to be sent per tier.
  • Speed range: our shim allows up to 4×, but Chirp 3 and Gemini cap at 2× and will reject higher.
  • Formats: 3.8 refuses MP3 and ignores the sample rate (always 24 kHz), so the shim keeps converting.
  • Cache: the audio cache is keyed on text + voice + speed + format. Every new setting (style, markup, pronunciations, gain…) must join that key, or a whispered reply could replay as a normal one.
  • Size: 5,000 bytes per request on Cloud TTS (Gemini: 4,000 for text plus 4,000 for the style prompt).

What I'd build first

(a) 3.8 style + tags, with the transcript stripping them. (b) A Pace setting that uses style on 3.8 and the native speed elsewhere. (c) Log the billed tokens 3.8 reports into the new usage page, so Gemini voice costs are exact rather than estimated. Streaming and a designed voice come after you've rated the clips.

Full field list (reference)

Per tier: live we tried it and it worked · doc documented, not tried · rejected Google refused it.

Cloud TTS text:synthesize (v1, GA)

POST https://texttospeech.googleapis.com/v1/text:synthesize · docs

FieldWhat it doesApp sends it todayStandardWaveNetNeural2StudioChirp HDChirp 3 HDGemini 2.5/3.1Gemini 3.8
input.text
string
The raw text to be synthesized.
Plain text. Whole request limit 5000 bytes (Gemini: text<=4000 B and prompt<=4000 B, sum<=8000 B). Exactly one of text/ssml/markup/multiSpeakerMarkup.
yesliveliveliveliveliveliveliven/a
input.ssml
string
The SSML document to be synthesized. The SSML document must be valid and well-formed. Otherwise the RPC will fail and return google.rpc.Code.INVALID_ARGUMENT. For more information, see [SSML](https://cloud.google.com/text-to-speech/docs/ssml).
LIVE: Chirp-HD -> HTTP 400; Chirp 3 HD honors <break> (preview, sync only, not in streaming). Studio drops <mark>,<emphasis>,<prosody pitch>,<lang>. Shim sends text, so SSML would be read literally. Gemini: no SSML documented.
nodocliveliveliverejectedliveunknownn/a
input.markup
string
Markup for Chirp 3: HD voices specifically. This field may not be used with any other voices.
Chirp 3 HD only: pause tags [pause], [pause short], [pause long] work ONLY here, not in text (live: markup 3.98 s vs text 4.54 s, text path appears to read tag aloud).
norejectedrejectedrejectedrejectedrejectedliverejectedn/a
input.prompt
string
This system instruction is supported only for controllable/promptable voice models. If this system instruction is used, we pass the unedited text to Gemini-TTS. Otherwise, a default system instruction is used. AI Studio calls this system instruction, Style Instructions.
Gemini style instruction, <=4000 bytes. Not used by 3.8 (use speechMetadata.style).
non/an/an/an/an/an/aliven/a
input.customPronunciations.pronunciations[]
array<CustomPronunciationParams>
The pronunciation customizations are applied.
Per-phrase IPA/X-SAMPA/Yomigana/Pinyin override; en-US first, many locales excluded for Chirp3 (see chirp3-hd doc). Instant-clone voices unsupported. Phrase must match exactly; use unique tokens (read1/read2) to disambiguate.
nounknownunknownunknownunknownunknownliveunknownn/a
input.customPronunciations.pronunciations[].phrase
string
The phrase to which the customization is applied. The phrase can be multiple words, such as proper nouns, but shouldn't span the length of the sentence.nounknownunknownunknownunknownunknownliveunknownn/a
input.customPronunciations.pronunciations[].phoneticEncoding
string
The phonetic encoding of the phrase.
Values: PHONETIC_ENCODING_UNSPECIFIED, PHONETIC_ENCODING_IPA, PHONETIC_ENCODING_X_SAMPA, PHONETIC_ENCODING_JAPANESE_YOMIGANA, PHONETIC_ENCODING_PINYIN
nounknownunknownunknownunknownunknownliveunknownn/a
input.customPronunciations.pronunciations[].pronunciation
string
The pronunciation of the phrase. This must be in the phonetic encoding specified above.nounknownunknownunknownunknownunknownliveunknownn/a
input.multiSpeakerMarkup.turns[]
array<Turn>
Required. Speaker turns.
Structured dialogue: turns[{speaker,text}]; speakers must match voice.multiSpeakerVoiceConfig aliases (Gemini) or Studio multispeaker names (e.g. R,S,T,U).
non/an/an/adocn/an/adocn/a
voice.languageCode
string
Required. The language (and potentially also the region) of the voice expressed as a [BCP-47](https://www.rfc-editor.org/rfc/bcp/bcp47.txt) language tag, e.g. "en-US". This should not include a script tag (e.g. use "cmn-cn" rather than "cmn-Hant-cn"), because the script will be inferred from the input provided in the SynthesisInput. The TTS service will use this parameter to help choose an appropriate voice. Note tha
Shim derives it from the voice name (first two hyphen segments); Gemini uses DEFAULT_LANGUAGE en-US.
yesliveliveliveliveliveliveliven/a
voice.name
string
The name of the voice. If both the name and the gender are not set, the service will choose a voice based on the other parameters such as language_code.
Classic: en-US-Wavenet-D etc. Chirp3: en-US-Chirp3-HD-Kore. Gemini: bare speaker (Kore) + modelName.
yesliveliveliveliveliveliveliven/a
voice.modelName
string
Optional. The name of the model. If set, the service will choose the model matching the specified configuration.
Gemini model id: gemini-2.5-flash-tts, gemini-2.5-flash-lite-preview-tts, gemini-2.5-pro-tts, gemini-3.1-flash-tts-preview.
yesn/an/an/an/an/an/aliven/a
voice.ssmlGender
string
The preferred gender of the voice. If not set, the service will choose a voice based on the other parameters such as language_code and name. Note that this is only a preference, not requirement; if a voice of the appropriate gender is not available, the synthesizer should substitute a voice with a different gender rather than failing the request.
Values: SSML_VOICE_GENDER_UNSPECIFIED, MALE, FEMALE, NEUTRAL
Preference only; ignored when name is set. NEUTRAL not yet supported.
nodocdocdocdocdocdocdocn/a
voice.customVoice.model
string
Required. The name of the AutoML model that synthesizes the custom voice.
AutoML Custom Voice resource; not in our tiers.
non/an/an/an/an/an/an/an/a
voice.customVoice.reportedUsage
string
Optional. Deprecated. The usage of the synthesized audio to be reported.
Values: REPORTED_USAGE_UNSPECIFIED, REALTIME, OFFLINE
Deprecated.
non/an/an/an/an/an/an/an/a
voice.voiceClone.voiceCloningKey
string
Required. Created by GenerateVoiceCloningKey.
Chirp 3 HD Instant Custom Voice key (GenerateVoiceCloningKey, allowlisted access). Not tested. 3.8 has its own voice replication.
non/an/an/an/an/aunknownn/an/a
voice.multiSpeakerVoiceConfig.speakerVoiceConfigs[]
array<MultispeakerPrebuiltVoice>
Required. A list of configurations for the voices of the speakers. Exactly two speaker voice configurations must be provided.
Exactly two [{speakerAlias,speakerId}].
non/an/an/an/an/an/adocn/a
audioConfig.audioEncoding
string
Required. The format of the audio byte stream.
Values: AUDIO_ENCODING_UNSPECIFIED, LINEAR16, MP3, MP3_64_KBPS, OGG_OPUS, MULAW, ALAW, PCM, M4A
Shim sends MP3, LINEAR16, OGG_OPUS. LIVE: v1 rejected MP3_64_KBPS (400, enum invalid in v1); M4A returned HTTP 200 on WaveNet (decode not checked). Chirp3: ALAW unsupported (doc); streaming supports only PCM/ALAW/MULAW/OGG_OPUS. Gemini unary: LINEAR16,ALAW,MULAW,MP3,OGG_OPUS,PCM. Enum descriptions: MP3=32kbps, MP3_64_KBPS=64kbps, LINEAR16 has WAV header, PCM headerless, MULAW/ALAW 8-bit G.711 with WAV header.
yesliveliveliveliveliveliveliven/a
audioConfig.speakingRate
number(double)
Optional. Input only. Speaking rate/speed, in the range [0.25, 2.0]. 1.0 is the normal native speed supported by the specific voice. 2.0 is twice as fast, and 0.5 is half as fast. If unset(0.0), defaults to the native 1.0 speed. Any other values < 0.25 or > 2.0 will return an error.
Values: 1.0
Default: Range 0.25-4.0 for classic voices per older docs; discovery text says [0.25, 2.0] (cap) while the shim clamps to 4.0. Chirp3 pace 0.25-2.0. LIVE: Chirp-HD 3.86->2.69 s @1.5, Chirp3 4.49->2.71, Gemini2.5 3.98->2.64. Shim clamps 0.25-4.0 - clamp to 2.0 for Chirp3/Gemini.
yesdoclivedocdocliveliveliven/a
audioConfig.pitch
number(double)
Optional. Input only. Speaking pitch, in the range [-20.0, 20.0]. 20 means increase 20 semitones from the original pitch. -20 means decrease 20 semitones from the original pitch.
Semitones -20..+20. Chirp3 -> HTTP 400 'does not support pitch'. LIVE accepted (HTTP 200) on Gemini 2.5 and Studio but audible effect not verified.
nodocdocdocdocunknownrejectedliven/a
audioConfig.volumeGainDb
number(double)
Optional. Input only. Volume gain (in dB) of the normal native volume supported by the specific voice, in the range [-96.0, 16.0]. If unset, or set to a value of 0.0 (dB), will play at normal native signal amplitude. A value of -6.0 (dB) will play at approximately half the amplitude of the normal native signal amplitude. A value of +6.0 (dB) will play at approximately twice the amplitude of the normal native signal a
-96..+16 dB; >+10 no benefit. Accepted live on Chirp3 and Gemini2.5; effect not measured. Cheap loudness normalization for phone speaker.
nodoclivedocdocunknownliveliven/a
audioConfig.sampleRateHertz
integer(int32)
Optional. The synthesis sample rate (in hertz) for this audio. When this is specified in SynthesizeSpeechRequest, if this is different from the voice's natural sample rate, then the synthesizer will honor this request by converting to the desired sample rate (which might result in worse audio quality), unless the specified sample rate is not supported for the encoding chosen, in which case it will fail the request an
Resamples from natural rate (Chirp3 LINEAR16 16 kHz returned HTTP 200). Unsupported rate for the encoding -> 400.
nodocdocdocdocdocliveunknownn/a
audioConfig.effectsProfileId[]
array<string>
Optional. Input only. An identifier which selects 'audio effects' profiles that are applied on (post synthesized) text to speech. Effects are applied on top of each other in the order they are given. See [audio profiles](https://cloud.google.com/text-to-speech/docs/audio-profiles) for current supported profile ids.
Values: wearable-class-device, handset-class-device, headphone-class-device, small-bluetooth-speaker-class-device, medium-bluetooth-speaker-class-device, large-home-entertainment-class-device, large-automotive-class-device, telephony-class-application
Post-synthesis EQ chain applied in given order; do not repeat a profile.
nodoclivedocdocdocunknownunknownn/a
advancedVoiceOptions.enableTextnorm
boolean
Optional. If true, textnorm will be applied to text input. This feature is enabled by default. Only applies for Gemini TTS.
Default: true
Gemini only; text normalization (numbers/dates) on by default. LIVE: false accepted.
non/an/an/an/an/an/aliven/a
advancedVoiceOptions.safetySettings.settings[].category
string
The harm category to apply the safety setting to.
Values: HARM_CATEGORY_UNSPECIFIED, HARM_CATEGORY_HATE_SPEECH, HARM_CATEGORY_DANGEROUS_CONTENT, HARM_CATEGORY_HARASSMENT, HARM_CATEGORY_SEXUALLY_EXPLICIT
Gemini TTS only. If no settings, filters are OFF by default.
non/an/an/an/an/an/adocn/a
advancedVoiceOptions.safetySettings.settings[].threshold
string
The harm block threshold for the safety setting.
Values: HARM_BLOCK_THRESHOLD_UNSPECIFIED, BLOCK_LOW_AND_ABOVE, BLOCK_MEDIUM_AND_ABOVE, BLOCK_ONLY_HIGH, BLOCK_NONE, OFF
non/an/an/an/an/an/adocn/a
advancedVoiceOptions.relaxSafetyFilters
boolean
Optional. Input only. Deprecated, use safety_settings instead. If true, relaxes safety filters for Gemini TTS.
Deprecated -> use safetySettings.
non/an/an/an/an/an/adocn/a
advancedVoiceOptions.lowLatencyJourneySynthesis
boolean
Only for Journey voices. If false, the synthesis is context aware and has a higher latency.
Journey voices only; none in our tiers.
non/an/an/an/an/an/an/an/a
audioContent
string(byte)
Base64 audio in the requested encoding (LINEAR16/OGG/MP3 include container header). No usage/token info in the response.yesliveliveliveliveliveliveliven/a

Cloud TTS text:synthesize v1beta1 (extras over v1)

POST https://texttospeech.googleapis.com/v1beta1/text:synthesize · docs

FieldWhat it doesApp sends it todayStandardWaveNetNeural2StudioChirp HDChirp 3 HDGemini 2.5/3.1Gemini 3.8
enableTimePointing[]
array<string>
Whether and what timepoints are returned in the response.
Values: TIMEPOINT_TYPE_UNSPECIFIED, SSML_MARK
Only SSML_MARK: returns timepoints for <mark name=.../> tags (word/sentence sync, captions, orb). LIVE Neural2: marks at 0.37 s and 0.84 s. Chirp3: request accepted but timepoints empty. Studio drops <mark>. Requires /v1beta1 endpoint. [v1beta1 only]
nodocdocliverejectedunknownrejectedunknownn/a
timepoints[]
array<Timepoint{markName,timeSeconds}>
Mark times (v1beta1 only).nodocdoclivedocdocrejecteddocn/a
audioConfig (response)
AudioConfig
Echo of the audio metadata (v1beta1 only).nodocdocdocdocdocdocdocn/a

Cloud TTS StreamingSynthesize (bidirectional gRPC; no REST/discovery)

gRPC google.cloud.texttospeech.v1.TextToSpeech/StreamingSynthesize (global/us/eu endpoints) · docs

FieldWhat it doesApp sends it todayStandardWaveNetNeural2StudioChirp HDChirp 3 HDGemini 2.5/3.1Gemini 3.8
streaming_config.voice
VoiceSelectionParams
Required; same as unary voice.non/an/an/an/an/adocdocn/a
streaming_config.streaming_audio_config.audio_encoding
enum
Required. All other encodings error.
Values: PCM, ALAW, MULAW, OGG_OPUS
Chirp3 ALAW unsupported per voice docs.
non/an/an/an/an/adocdocn/a
streaming_config.streaming_audio_config.sample_rate_hertz
int32
Optional output rate.non/an/an/an/an/adocdocn/a
streaming_config.streaming_audio_config.speaking_rate
double
Pace; no pitch/volume in streaming config.
Values: 0.25-2.0
Default: 1.0
non/an/an/an/an/adocdocn/a
streaming_config.custom_pronunciations
CustomPronunciations
Same as unary; en-US first.non/an/an/an/an/adocn/an/a
streaming_config.advanced_voice_options
AdvancedVoiceOptions
Gemini safety/textnorm.non/an/an/an/an/an/adocn/a
input.text
string
Chunk of text; send complete sentences for better prosody.non/an/an/an/an/adocdocn/a
input.markup
string
Chirp 3 HD pause markup.non/an/an/an/an/adocn/an/a
input.multi_speaker_markup
MultiSpeakerMarkup
Gemini multi-speaker.non/an/an/an/an/an/adocn/a
input.prompt
string
Gemini style instruction (in the first input chunk).non/an/an/an/an/an/adocn/a
response.audio_content
bytes
Headerless LINEAR16 24 kHz chunks.non/an/an/an/an/adocdocn/a

Cloud TTS synthesizeLongAudio (async, GCS output)

POST https://texttospeech.googleapis.com/v1/{parent=projects/*/locations/*}:synthesizeLongAudio ; poll GET v1/{name=projects/*/locations/*/operations/*} · docs

FieldWhat it doesApp sends it todayStandardWaveNetNeural2StudioChirp HDChirp 3 HDGemini 2.5/3.1Gemini 3.8
parent (path)
string
projects/*/locations/* (docs example uses locations/global on v1beta1 and us-central1).nounknownunknownunknownunknownunknownunknownunknownn/a
input
SynthesisInput
Text or SSML, up to ~1 MB.nounknownunknownunknownunknownunknownunknownunknownn/a
voice
object
Required. The desired voice of the synthesized audio.nounknownunknownunknownunknownunknownunknownunknownn/a
audioConfig
object
Required. The configuration of the synthesized audio.
Docs example LINEAR16.
nounknownunknownunknownunknownunknownunknownunknownn/a
outputGcsUri
string
Required. Specifies a Cloud Storage URI for the synthesis results. Must be specified in the format: `gs://bucket_name/object_name`, and the bucket must already exist.
gs://bucket/object; bucket must exist.
nounknownunknownunknownunknownunknownunknownunknownn/a
operations.get/list/cancel/delete
REST
Long-running operation management (poll progressPercentage in metadata).nounknownunknownunknownunknownunknownunknownunknownn/a

Cloud TTS voices.list

GET https://texttospeech.googleapis.com/v1/voices?languageCode=en-US · docs

FieldWhat it doesApp sends it todayStandardWaveNetNeural2StudioChirp HDChirp 3 HDGemini 2.5/3.1Gemini 3.8
languageCode (query)
string
BCP-47 filter; 'no' also returns nb-*.
Default: all voices
yesdocdocdocdocdocdocdocn/a
voices[].name
string
The name of this voice. Each distinct voice has a unique name.yesdocdocdocdocdocdocdocn/a
voices[].languageCodes[]
array<string>
The languages that this voice supports, expressed as [BCP-47](https://www.rfc-editor.org/rfc/bcp/bcp47.txt) language tags (e.g. "en-US", "es-419", "cmn-tw").yesdocdocdocdocdocdocdocn/a
voices[].ssmlGender
string
The gender of this voice.
Values: SSML_VOICE_GENDER_UNSPECIFIED, MALE, FEMALE, NEUTRAL
nodocdocdocdocdocdocdocn/a
voices[].naturalSampleRateHertz
integer(int32)
The natural sample rate (in hertz) for this voice.
Chirp3/Gemini 24000; classic 24000 (WaveNet/Neural2) or 22050 (Standard).
nodocdocdocdocdocdocdocn/a

Gemini Enterprise Voices API (3.8 voice design / replication / library)

POST/GET/DELETE https://aiplatform.googleapis.com/v1beta1/projects/{p}/locations/global/voices[:create|:list|:get|:delete] · docs

FieldWhat it doesApp sends it todayStandardWaveNetNeural2StudioChirp HDChirp 3 HDGemini 2.5/3.1Gemini 3.8
store
boolean
true: Google stores and returns voice_<id> (1 yr after last use, per-project quota). false: returns client-held voicekey_ (7 days). prompted voices REQUIRE store=true.
Values: True, False
Default: false
non/an/an/an/an/an/an/adoc
voice.prompted
object
Parameters for prompted voice generation. Required in `CreateVoice` when `type` is `"prompted"`. Returned in `CreateVoice`, `GetVoice`, and `ListVoices` responses for prompted voices.non/an/an/an/an/an/an/adoc
voice.sampleAudio
object
Output only. Sample audio (synthesizer prompt audio) generated for a prompted voice. Populated only in `CreateVoice` and `GetVoice` responses when `type` is `"prompted"`; unset in `ListVoices` responses and for `"replicated"` or `"prebuilt"` voices.non/an/an/an/an/an/an/adoc
voice.type
string
Required. The type of the voice. In `CreateVoice`, must be `"replicated"` or `"prompted"`. In `ListVoices` responses, may also be `"prebuilt"`.
Values: VOICE_TYPE_UNSPECIFIED, VOICE_TYPE_REPLICATED, VOICE_TYPE_PROMPTED, VOICE_TYPE_PREBUILT
non/an/an/an/an/an/an/adoc
voice.description
string
Optional. Descriptive summary of vocal timbre, personality, and tone.non/an/an/an/an/an/an/adoc
voice.pitch
string
Optional. Voice pitch classification.
Values: PITCH_UNSPECIFIED, PITCH_LOW, PITCH_MEDIUM, PITCH_HIGH
non/an/an/an/an/an/an/adoc
voice.expireTime
string(google-datetime)
Output only. The timestamp at which a custom stored voice (`store = true`) or replicated voice key (`store = false`) expires. For custom stored voices (`store = true`), this expiration time is extended when the voice is used for speech synthesis or as a `base_voice` in `CreateVoice`. Unset for prebuilt catalog voices (`"prebuilt"`), which do not expire.non/an/an/an/an/an/an/adoc
voice.usage
object
Output only. Token usage statistics for the voice creation request. Populated only in the response of `CreateVoice` when `type` is `"prompted"`; unset for `"replicated"` and in `GetVoice` and `ListVoices` responses.non/an/an/an/an/an/an/adoc
voice.persona
string
Optional. Intended persona or character archetype (e.g. "Warm, Friendly", "Narrator").non/an/an/an/an/an/an/adoc
voice.accent
string
Optional. Regional accent descriptor (e.g. "American", "British").non/an/an/an/an/an/an/adoc
voice.model
string
Optional. The model used to design or replicate the voice. If omitted in `CreateVoice`, defaults to the latest supported voice design model. Returned in `CreateVoice`, `GetVoice`, and `ListVoices` responses for custom voices (`"prompted"` and `"replicated"`); unset for `"prebuilt"` voices. Created voices can be synthesized across any supported TTS synthesis non/an/an/an/an/an/an/adoc
voice.languageCode
string
Optional. Primary BCP-47 language tag (e.g. "en-US", "fr-FR").non/an/an/an/an/an/an/adoc
voice.gender
string
Optional. Perceived voice gender presentation (e.g. "female", "male", "neutral").non/an/an/an/an/an/an/adoc
voice.replicated
object
Input only. Parameters for replicated voice generation. Required on input when `type` is `"replicated"`. Not returned in responses.non/an/an/an/an/an/an/adoc
voice.key
string
Output only. The client-managed voice replication key (with prefix `voicekey_`). Returned only by `CreateVoice` when `type` is `"replicated"` and `store` is `false`. Pass this key to `SpeechConfig.voice_config.voice` (or `SpeechConfig.voice`) during speech synthesis.non/an/an/an/an/an/an/adoc
voice.id
string
Output only. The unique identifier of the voice. * For Google-managed custom voices (`store = true`), this is a generated ID with prefix `voice_` (for example, `voice_abc123def456`). Pass `voices/{id}` as the `name` in `GetVoice` and `DeleteVoice`, and pass `{id}` directly to `SpeechConfig.voice_config.voice` (or `SpeechConfig.voice`) during speech synthesisnon/an/an/an/an/an/an/adoc
voice.regionCode
string
Optional. ISO 3166-1 alpha-2 or UN M.49 geographic region code (e.g. "US", "GB", "001").non/an/an/an/an/an/an/adoc
voice.displayName
string
Optional. User-provided display name for a stored voice (`store = true`), or the catalog name for a prebuilt voice.non/an/an/an/an/an/an/adoc
voice.context
string
Optional. Optimal usage context or domain (e.g. "Conversational", "Audiobook", "News").non/an/an/an/an/an/an/adoc
voice.prompted.input
string
Required for VOICE_TYPE_PROMPTED: natural-language persona (age, timbre, accent, delivery). Designing takes several seconds (60 s timeout).
LIVE: create+synthesize+delete worked with store=true, display_name, gender, language_code, prompted.input
non/an/an/an/an/an/an/adoc
voice.replicated.sourceAudio / consentAudio
AudioData
10-30 s adult reference + consent-phrase recording.
DOC only; not run.
non/an/an/an/an/an/an/adoc
list query: pageToken,pageSize(<=50),search,type(prebuilt|prompted|replicated),gender,pitch,persona,context,accent,regionCode,languageCode
query
Browse 2000+ Extended Library voices + own voices. Doc says language_code/gender filters fail with INVALID_ARGUMENT; discovery lists them.
LIVE: search=narrator and search='conversational agent' worked.
non/an/an/an/an/an/an/adoc

Gemini Enterprise generateContent / streamGenerateContent (Gemini 3.8 TTS; also serves 2.5/3.1 on Vertex)

POST https://aiplatform.googleapis.com/v1/projects/{p}/locations/global/publishers/google/models/{model}:generateContent · docs

FieldWhat it doesApp sends it todayStandardWaveNetNeural2StudioChirp HDChirp 3 HDGemini 2.5/3.1Gemini 3.8
contents[].role
string
Turn role.
Values: user
non/an/an/an/an/an/an/alive
contents[].parts[].text
string
Verbatim transcript. Instructions in text may be spoken aloud. Angle-bracket tags: <sigh> <laugh> <cough> <gasp> <breath> <short pause> <long pause> <whispers> and ~35 more. CAPS = emphasis. |word| = backchannel/overlap (multi-speaker).
Shim sends text only. Gemini 2.5/3.1 on Vertex take style/tags inside the text itself, with [tag] syntax.
yesn/an/an/an/an/an/adoclive
contents[].parts[].speechMetadata.style
string
Turn-level delivery: emotion, pace, volume (e.g. 'whispering softly', 'speaking rapidly'). Keep short and constant; 'speaking rapidly' cut audio 33% live.
Default: empty
Billed as input text tokens (+2 to +4 tokens).
non/an/an/an/an/an/an/alive
contents[].parts[].speechMetadata.speaker
string
Speaker name per turn; must match a speakerVoiceConfigs speaker.non/an/an/an/an/an/an/alive
generationConfig.responseModalities[]
array<string>
Must include AUDIO.
Values: AUDIO
yesn/an/an/an/an/an/an/alive
generationConfig.speechConfig.voiceConfig.voice
string
Prebuilt speaker (Kore), Extended Library id, designed voice_..., or voicekey_...
Shim forwards the Speaker name only.
yesn/an/an/an/an/an/an/alive
generationConfig.speechConfig.voiceConfig.prebuiltVoiceConfig.voiceName
string
Older spelling; still works for prebuilt.non/an/an/an/an/an/an/adoc
generationConfig.speechConfig.voiceConfig.replicatedVoiceConfig
object
Inline reference audio; NOT supported on 3.8 (use Voices API).non/an/an/an/an/an/an/arejected
generationConfig.speechConfig.languageCode
string
Optional hint; 3.8 auto-detects (Flash 130 languages, Flash-Lite 101).
Default: auto-detected
non/an/an/an/an/an/an/adoc
generationConfig.speechConfig.multiSpeakerVoiceConfig.speakerVoiceConfigs[]
array
Exactly two {speaker, voiceConfig.voice}. Mutually exclusive with voiceConfig.non/an/an/an/an/an/an/alive
generationConfig.speechConfig.speakingRate
double
Does not exist in 3.8 schema.
LIVE: HTTP 400 Unknown name speakingRate.
non/an/an/an/an/an/an/arejected
generationConfig.responseFormat[].audio.mimeType
enum
Output encoding. Schema lists MP3 and OGG_OPUS but docs say unsupported: LIVE AUDIO_MP3 (+bitRate) -> HTTP 400. MULAW/ALAW come back 8 kHz raw (LIVE MULAW ok, no WAV header).
Values: AUDIO_WAV, AUDIO_L16, AUDIO_MULAW, AUDIO_ALAW, AUDIO_MP3, AUDIO_OGG_OPUS
Default: AUDIO_WAV unary / AUDIO_L16 streaming
Shim transcodes WAV->MP3 itself.
non/an/an/an/an/an/an/alive
generationConfig.responseFormat[].audio.sampleRate
int32
Ignored by 3.8: LIVE sampleRate=16000 still returned a 24 kHz WAV.
Default: 24000
non/an/an/an/an/an/an/arejected
generationConfig.responseFormat[].audio.bitRate
int32
Compressed formats only; moot since MP3/Opus unsupported.non/an/an/an/an/an/an/arejected
generationConfig.responseFormat[].audio.delivery
enum
Inline base64 vs URI delivery.
Values: INLINE, URI, DELIVERY_UNSPECIFIED
Default: INLINE
non/an/an/an/an/an/an/aunknown
generationConfig.temperature / topP / topK / candidateCount / seed / stopSequences ...
various
Rejected with INVALID_ARGUMENT on 3.8 (doc). Ignored on 2.5/3.1.non/an/an/an/an/an/adocrejected
systemInstruction
Content
Rejected on 3.8; put persona in Voices API instead.non/an/an/an/an/an/an/arejected
response.candidates[0].content.parts[0].inlineData{data,mimeType}
object
Base64 WAV (unary). Shim decodes WAV or audio/L16.yesn/an/an/an/an/an/an/alive
response.usageMetadata.{promptTokenCount,candidatesTokenCount,candidatesTokensDetails[AUDIO],promptTokensDetails[TEXT]}
object
Billing basis. Audio output = 25 tokens per second of audio (measured 25.0). LIVE present on 3.8; absent on Cloud TTS.
Shim could log it for exact cost.
non/an/an/an/an/an/an/alive
streamGenerateContent?alt=sse
method
Streams raw L16 24 kHz chunks; first-audio latency win.non/an/an/an/an/an/an/adoc
0 / 93