feat: Gemini 3.8 TTS dialect + corpus enforcement (0.5.0)

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "speechmarkdown-rust"
-version = "0.4.13"
+version = "0.5.0"
 edition = "2021"
 authors = ["SpeechMarkdown Contributors"]
 description = "High-performance SpeechMarkdown parser with multi-language bindings"

build.rs

@@ -1,4 +1,4 @@
 fn main() {
     // Tell cargo to rerun this build script if important files change
     println!("cargo:rerun-if-changed=src/");
-}
+}

examples/gen_gemini_fixtures.rs

@@ -0,0 +1,56 @@
+//! Generates `.gemini.ssml` corpus expectations for the shared fixture
+//! inputs, in the speechmarkdown-test-files layout.
+//!
+//! Run from the repo root:
+//!   cargo run --example gen_gemini_fixtures -- <output-dir>
+//!
+//! Inputs mirror the elevenlabs-v3-* set where one exists (so dialects
+//! are directly comparable); the rest cover gemini-specific policies
+//! (backchannels as plain text, sound effects dropped, pause thresholds).
+use speechmarkdown_rust::{Platform, SpeechMarkdownParser};
+use std::fs;
+use std::path::Path;
+
+fn main() {
+    let out_dir = std::env::args()
+        .nth(1)
+        .expect("usage: gen_gemini_fixtures <output-dir>");
+    let cases: Vec<(&str, &str)> = vec![
+        ("gemini-break-short", "Sample [3s] speech [250ms] markdown"),
+        (
+            "gemini-break-strength",
+            "None [break:\"none\"] x-weak [break:\"x-weak\"]\nweak [break:\"weak\"] medium [break:\"medium\"]\nstrong [break:\"strong\"] x-strong [break:\"x-strong\"]",
+        ),
+        (
+            "gemini-emphasis-short",
+            "This is ++important++ and +small+ and -muted- and ~flat~.",
+        ),
+        ("gemini-expressive", "He [laugh] and then [applause] left."),
+        ("gemini-ipa-short", "Say (speech)/spitʃ/ clearly."),
+        (
+            "gemini-modifiers",
+            "A (secret)[whisper], a (fast thing)[rate:\"fast\"], and a (loud thing)[volume:\"loud\"].",
+        ),
+        ("gemini-section-style", "#[whisper] Hello world"),
+        ("gemini-sub-short", "The {AL}aluminum is recycled."),
+        (
+            "gemini-vocal-aliases",
+            "[cheering] Go team! [whew] ... [throat-clear] anyway.",
+        ),
+        ("gemini-backchannels", "I [mhm] agree. [oh] really? Well [wow]."),
+        (
+            "gemini-pauses",
+            "Hold on [750ms] alright... got it [1s] moving on [120ms] now.",
+        ),
+    ];
+
+    for (name, input) in cases {
+        let dir = Path::new(&out_dir).join(name);
+        fs::create_dir_all(&dir).expect("create dir");
+        fs::write(dir.join(format!("{name}.smd")), input).expect("write smd");
+        let out = SpeechMarkdownParser::to_ssml(input, Platform::Gemini)
+            .unwrap_or_else(|e| panic!("{name}: {e}"));
+        fs::write(dir.join(format!("{name}.gemini.ssml")), out).expect("write expected");
+        println!("wrote {name}");
+    }
+}

src/capabilities.rs

@@ -27,6 +27,7 @@ pub fn get_supported_ssml(platform: Platform) -> PlatformCapabilities {
         Platform::SamsungBixby => samsung_bixby_capabilities(),
         Platform::ElevenLabs => elevenlabs_capabilities(),
         Platform::ElevenLabsV3 => elevenlabs_v3_capabilities(),
+        Platform::Gemini => gemini_capabilities(),
         Platform::IbmWatson => ibm_watson_capabilities(),
     }
 }
@@ -529,6 +530,93 @@ fn elevenlabs_v3_capabilities() -> PlatformCapabilities {
     }
 }
 
+fn gemini_capabilities() -> PlatformCapabilities {
+    // Gemini 3.8 TTS dialect: no SSML. The transcript is a verbatim script
+    // directed with angle-bracket vocal bursts/pauses, CAPS emphasis and
+    // plain-text disfluencies; sustained delivery lives in the turn-level
+    // speech_metadata.style field (engine channel), not inline.
+    PlatformCapabilities {
+        platform: "gemini".into(),
+        ssml_elements: vec![
+            SsmlCapability {
+                element: "<vocal burst>".into(),
+                description: "Angle-bracket vocal bursts (<laugh>, <sigh>, \
+                              <gasp>, <cough>, <chuckle>, <giggle>, <groan>, \
+                              <scream>, <yawn>, <throat-clearing>, ...); \
+                              English tags recommended even for non-English \
+                              transcripts"
+                    .into(),
+                attributes: vec![],
+                speech_markdown_syntax: vec!["[laugh]".into(), "[sigh]".into(), "[gasp]".into()],
+                example: "Wait... did you hear that? <sigh>".into(),
+            },
+            SsmlCapability {
+                element: "<short pause>/<long pause>".into(),
+                description: "Two pause steps only; sub-beat pauses become \
+                              ellipses (...). No exact durations"
+                    .into(),
+                attributes: vec![],
+                speech_markdown_syntax: vec![
+                    "[500ms]".into(),
+                    "[2s]".into(),
+                    "[break:strong]".into(),
+                ],
+                example: "Hold on, let me think... <short pause> Alright".into(),
+            },
+            SsmlCapability {
+                element: "CAPS emphasis".into(),
+                description: "Emphasis via capitalization (moderate and \
+                              strong collapse to one level)"
+                    .into(),
+                attributes: vec![],
+                speech_markdown_syntax: vec!["++word++".into(), "+word+".into()],
+                example: "This is a VERY important point!".into(),
+            },
+            SsmlCapability {
+                element: "<whispers>".into(),
+                description: "Whisper as an inline vocal burst".into(),
+                attributes: vec![],
+                speech_markdown_syntax: vec!["(text)[whisper]".into(), "#[whisper] text".into()],
+                example: "<whispers> it's a secret".into(),
+            },
+            SsmlCapability {
+                element: "plain-text disfluencies".into(),
+                description: "Backchannels ([mhm], [oh], [wow], ...) are \
+                              spoken as plain text — Gemini wants transcripts \
+                              written like real speech"
+                    .into(),
+                attributes: vec![],
+                speech_markdown_syntax: vec!["[mhm]".into(), "[oh]".into(), "[yeah]".into()],
+                example: "Oh uh yeah I think... hm, so that's interesting".into(),
+            },
+        ],
+        unsupported: vec![
+            "break (no SSML; two-tag pause approximation)".into(),
+            "prosody (rate/pitch/volume are turn-level speech_metadata.style, \
+             an engine channel — dropped inline)"
+                .into(),
+            "phoneme (no equivalent; Gemini handles pronunciation natively, \
+             the word itself is spoken)"
+                .into(),
+            "say-as (text normalization is native)".into(),
+            "sub (alias is spoken instead of the text)".into(),
+            "audio (no equivalent)".into(),
+            "mark (no word timestamps from the API)".into(),
+            "voice (switch via generationConfig.speech_config)".into(),
+            "lang (no equivalent)".into(),
+            "[applause]/[boo] and other sound effects (prompting guide says \
+             avoid non-vocal SFX)"
+                .into(),
+            "emphasis levels (moderate/strong both become CAPS)".into(),
+            "amazon:effect".into(),
+            "amazon:emotion".into(),
+            "amazon:domain".into(),
+            "mstts:express-as".into(),
+            "google:style".into(),
+        ],
+    }
+}
+
 fn ibm_watson_capabilities() -> PlatformCapabilities {
     PlatformCapabilities {
         platform: "ibm-watson".into(),

src/formatters/base.rs

@@ -15,6 +15,10 @@ pub enum Platform {
     /// Eleven v3 audio-tag dialect (eleven_v3 / eleven_v3_conversational):
     /// bracketed natural-language tags instead of SSML.
     ElevenLabsV3,
+    /// Gemini 3.8 TTS dialect (gemini-3.8-flash-tts / -flash-lite-tts):
+    /// angle-bracket vocal bursts and pauses, CAPS emphasis, and plain-text
+    /// disfluencies instead of SSML.
+    Gemini,
     IbmWatson,
 }
 
@@ -30,6 +34,7 @@ impl Platform {
             "samsung-bixby" | "bixby" => Some(Platform::SamsungBixby),
             "elevenlabs" => Some(Platform::ElevenLabs),
             "elevenlabs-v3" | "elevenlabs_v3" | "eleven-v3" => Some(Platform::ElevenLabsV3),
+            "gemini" | "gemini-3.8" | "gemini-tts" | "google-gemini" => Some(Platform::Gemini),
             "ibm-watson" | "watson" => Some(Platform::IbmWatson),
             _ => None,
         }
@@ -46,6 +51,7 @@ impl Platform {
             Platform::SamsungBixby => "samsung-bixby",
             Platform::ElevenLabs => "elevenlabs",
             Platform::ElevenLabsV3 => "elevenlabs-v3",
+            Platform::Gemini => "gemini",
             Platform::IbmWatson => "ibm-watson",
         }
     }
@@ -120,6 +126,12 @@ pub fn create_formatter(platform: Platform, options: FormatterOptions) -> Box<dy
             // natural-language tags, punctuation pauses, native slash IPA.
             Box::new(super::elevenlabs_v3::ElevenLabsV3Formatter::new(options))
         }
+        Platform::Gemini => {
+            // Gemini 3.8 TTS dialect: no SSML — the transcript is directed
+            // with angle-bracket vocal bursts/pauses, CAPS emphasis, and
+            // plain-text disfluencies.
+            Box::new(super::gemini::GeminiFormatter::new(options))
+        }
         _ => Box::new(super::TextFormatter::new()),
     }
 }