AACTools · willwade · Sep 4, 2026 · Sep 4, 2026 · Sep 4, 2026 · Sep 4, 2026
diff --git a/README.md b/README.md
@@ -48,9 +48,24 @@ You can safely ignore Swift-specific files. Your respective package managers (Ca
 | Apple | `"apple"` |
 | W3C | `"w3c"` |
 | Samsung Bixby | `"samsung-bixby"` or `"bixby"` |
-| ElevenLabs | `"elevenlabs"` |
+| ElevenLabs (pre-v3) | `"elevenlabs"` |
+| ElevenLabs v3 | `"elevenlabs-v3"` |
 | IBM Watson | `"ibm-watson"` or `"watson"` |
 
+### ElevenLabs note
+
+ElevenLabs does not parse SSML documents, so both ElevenLabs platforms emit
+prompt markup rather than SSML, and the correct dialect depends on the model:
+
+- `"elevenlabs"` — pre-v3 models (`eleven_multilingual_v2`, `eleven_flash_v2_5`,
+  `eleven_flash_v2`, `eleven_turbo_v2`): `<break time="x.xs"/>` pauses (max 3s)
+  and `<phoneme>` for IPA (flash/turbo, English only). No `<speak>` wrapper.
+- `"elevenlabs-v3"` — `eleven_v3` / `eleven_v3_conversational`: bracketed
+  natural-language audio tags (`[whispers]`, `[excited]`, `[pause]`,
+  `[long pause]`, `[laughs]`), punctuation pauses, and native `"/IPA/"`.
+  Audio tags are model-interpreted direction — best-effort, voice-dependent —
+  and are not understood by pre-v3 models.
+
 ## API
 
 All bindings expose the same core methods:

diff --git a/bindings/dotnet/SpeechMarkdown.cs b/bindings/dotnet/SpeechMarkdown.cs
@@ -274,7 +274,9 @@ public static class Platform
         public const string Apple = "apple";
         public const string W3c = "w3c";
         public const string SamsungBixby = "samsung-bixby";
-        public const string ElevenLabs = "eleven-labs";
+        // Note: the core library parses "elevenlabs" (no hyphen).
+        public const string ElevenLabs = "elevenlabs";
+        public const string ElevenLabsV3 = "elevenlabs-v3";
         public const string IbmWatson = "ibm-watson";
     }
 }
diff --git a/bindings/nodejs/src/lib.rs b/bindings/nodejs/src/lib.rs
@@ -5,7 +5,7 @@ use speechmarkdown_rust::{Platform, SpeechMarkdownParser};
 fn parse_platform(platform: &str) -> Result<Platform> {
     Platform::from_platform_str(platform).ok_or_else(|| {
         Error::from_reason(format!(
-            "unsupported platform: '{}'. Use one of: amazon-alexa, google-assistant, microsoft-azure, apple, w3c, samsung-bixby, elevenlabs, ibm-watson",
+            "unsupported platform: '{}'. Use one of: amazon-alexa, google-assistant, microsoft-azure, apple, w3c, samsung-bixby, elevenlabs, elevenlabs-v3, ibm-watson",
             platform
         ))
     })

diff --git a/bindings/python/src/lib.rs b/bindings/python/src/lib.rs
@@ -4,7 +4,7 @@ use speechmarkdown_rust::{Platform, SpeechMarkdownParser};
 fn parse_platform(platform: &str) -> PyResult<Platform> {
     Platform::from_platform_str(platform).ok_or_else(|| {
         pyo3::exceptions::PyValueError::new_err(format!(
-            "unsupported platform: '{}'. Use one of: amazon-alexa, google-assistant, microsoft-azure, apple, w3c, samsung-bixby, elevenlabs, ibm-watson",
+            "unsupported platform: '{}'. Use one of: amazon-alexa, google-assistant, microsoft-azure, apple, w3c, samsung-bixby, elevenlabs, elevenlabs-v3, ibm-watson",
             platform
         ))
     })

diff --git a/bindings/swift/Sources/CSpeechMarkdown/include/speechmarkdown.h b/bindings/swift/Sources/CSpeechMarkdown/include/speechmarkdown.h
@@ -10,7 +10,7 @@ extern "C" {
 
 // Convert SpeechMarkdown input to SSML for the given platform.
 // Platforms: "amazon-alexa", "google-assistant", "microsoft-azure",
-//            "apple", "w3c", "samsung-bixby", "eleven-labs", "ibm-watson"
+//            "apple", "w3c", "samsung-bixby", "elevenlabs", "elevenlabs-v3", "ibm-watson"
 // Returns: allocated string with SSML, or NULL on error.
 //          Caller must free with speechmarkdown_free().
 const char* speechmarkdown_to_ssml(const char* input, const char* platform);

diff --git a/src/ast.rs b/src/ast.rs
@@ -162,6 +162,10 @@ pub enum NodeType {
     /// Mark tag
     Mark,
 
+    /// Expressive audio tag [laugh], [sigh], [applause], etc.
+    /// Kept verbatim by formatters as a bracketed audio tag.
+    Expressive,
+
     // Modifier types (for text modifiers and sections)
     /// Emphasis modifier
     Emphasis,

diff --git a/src/capabilities.rs b/src/capabilities.rs
@@ -26,6 +26,7 @@ pub fn get_supported_ssml(platform: Platform) -> PlatformCapabilities {
         Platform::W3c => w3c_capabilities(),
         Platform::SamsungBixby => samsung_bixby_capabilities(),
         Platform::ElevenLabs => elevenlabs_capabilities(),
+        Platform::ElevenLabsV3 => elevenlabs_v3_capabilities(),
         Platform::IbmWatson => ibm_watson_capabilities(),
     }
 }
@@ -396,24 +397,127 @@ fn samsung_bixby_capabilities() -> PlatformCapabilities {
 }
 
 fn elevenlabs_capabilities() -> PlatformCapabilities {
+    // Pre-v3 prompt markup (eleven_multilingual_v2 / flash_v2_5 /
+    // flash_v2 / turbo_v2): no SSML document, but two inline tags are
+    // parsed. Everything else degrades to plain text.
     PlatformCapabilities {
         platform: "elevenlabs".into(),
         ssml_elements: vec![
-            break_element(),
-            prosody_element(),
-            audio_element(),
-            phoneme_element(),
-            mark_element(),
-            say_as_element("characters", "characters"),
-            say_as_element("number", "number"),
-            say_as_element("date", "date"),
-            say_as_element("time", "time"),
+            SsmlCapability {
+                element: "break".into(),
+                description: "Insert a pause (pre-v3 models only; max 3s; \
+                              eleven_v3 does not parse break tags)"
+                    .into(),
+                attributes: vec!["time".into()],
+                speech_markdown_syntax: vec![
+                    "[2s]".into(),
+                    "[500ms]".into(),
+                    "[break:strong]".into(),
+                ],
+                example: "Hello [2s] world".into(),
+            },
+            SsmlCapability {
+                element: "phoneme".into(),
+                description: "Custom pronunciation (eleven_flash_v2 / \
+                              eleven_turbo_v2 only, English only)"
+                    .into(),
+                attributes: vec!["alphabet".into(), "ph".into()],
+                speech_markdown_syntax: vec![
+                    "(text)[ipa:\"pɪkəloʊ\"]".into(),
+                    "(text)/pɪkəloʊ/".into(),
+                ],
+                example: "(piccolo)/pɪkəloʊ/".into(),
+            },
         ],
         unsupported: vec![
-            "emphasis".into(),
-            "voice".into(),
-            "lang".into(),
-            "sub".into(),
+            "emphasis (no equivalent; degrades to text)".into(),
+            "prosody (rate maps to the API voice_settings.speed; use the \
+             elevenlabs-v3 dialect for tag-based control)"
+                .into(),
+            "say-as (text normalization is built in)".into(),
+            "sub (use pronunciation dictionaries or phonetic spelling)".into(),
+            "audio (no equivalent)".into(),
+            "mark (no equivalent)".into(),
+            "voice (switch via the API voice_id parameter)".into(),
+            "lang (no equivalent)".into(),
+            "audio tags (eleven_v3 only)".into(),
+            "amazon:effect".into(),
+            "amazon:emotion".into(),
+            "amazon:domain".into(),
+            "mstts:express-as".into(),
+            "google:style".into(),
+        ],
+    }
+}
+
+fn elevenlabs_v3_capabilities() -> PlatformCapabilities {
+    // Eleven v3 audio-tag dialect: no SSML at all. Bracketed
+    // natural-language tags are prompts interpreted by the model —
+    // best-effort, voice-dependent, and directional from their insertion
+    // point (no span semantics).
+    PlatformCapabilities {
+        platform: "elevenlabs-v3".into(),
+        ssml_elements: vec![
+            SsmlCapability {
+                element: "[audio tag]".into(),
+                description: "Natural-language performance direction in \
+                              brackets; open-ended (emotions, delivery, \
+                              reactions, sound effects, accents)"
+                    .into(),
+                attributes: vec![],
+                speech_markdown_syntax: vec![
+                    "[laugh]".into(),
+                    "#[excited] text".into(),
+                    "(text)[whisper]".into(),
+                ],
+                example: "[whispers] It's a secret".into(),
+            },
+            SsmlCapability {
+                element: "[pause] family".into(),
+                description: "Pauses via [short pause] / [pause] / [long \
+                              pause] or ellipses; approximate — no exact \
+                              durations on v3"
+                    .into(),
+                attributes: vec![],
+                speech_markdown_syntax: vec![
+                    "[500ms]".into(),
+                    "[2s]".into(),
+                    "[break:strong]".into(),
+                ],
+                example: "Hello [2s] world".into(),
+            },
+            SsmlCapability {
+                element: "inline IPA".into(),
+                description: "Native IPA wrapped in quotes and slashes".into(),
+                attributes: vec![],
+                speech_markdown_syntax: vec!["(speech)/spitʃ/".into()],
+                example: "(speech)/spitʃ/".into(),
+            },
+            SsmlCapability {
+                element: "emphasis tags".into(),
+                description: "Emphasis via [emphasized] / [stress on next \
+                              word] / [understated] (capitalization also \
+                              works but mutates the text)"
+                    .into(),
+                attributes: vec![],
+                speech_markdown_syntax: vec!["++word++".into(), "+word+".into()],
+                example: "++important++".into(),
+            },
+        ],
+        unsupported: vec![
+            "break (eleven_v3 does not parse SSML break tags)".into(),
+            "phoneme (use inline IPA)".into(),
+            "prosody (rate maps to the API voice_settings.speed; \
+             tempo/volume map to [rushed]/[drawn out]/[softly]/[loudly] tags)"
+                .into(),
+            "say-as (text normalization is built in)".into(),
+            "sub (alias is spoken instead of the text)".into(),
+            "audio (no equivalent)".into(),
+            "mark (no equivalent)".into(),
+            "voice (switch via the API voice_id parameter or Text to \
+             Dialogue)"
+                .into(),
+            "lang (no equivalent)".into(),
             "amazon:effect".into(),
             "amazon:emotion".into(),
             "amazon:domain".into(),
@@ -466,24 +570,64 @@ mod tests {
             Platform::W3c,
             Platform::SamsungBixby,
             Platform::ElevenLabs,
+            Platform::ElevenLabsV3,
             Platform::IbmWatson,
         ] {
             let caps = get_supported_ssml(platform);
-            assert!(!caps.ssml_elements.is_empty(), "{:?} has no elements", platform);
+            assert!(
+                !caps.ssml_elements.is_empty(),
+                "{:?} has no elements",
+                platform
+            );
             assert!(!caps.platform.is_empty());
         }
     }
 
+    #[test]
+    fn test_all_platforms_have_break() {
+        for platform in [
+            Platform::AmazonAlexa,
+            Platform::GoogleAssistant,
+            Platform::MicrosoftAzure,
+            Platform::Apple,
+            Platform::W3c,
+            Platform::SamsungBixby,
+            Platform::ElevenLabs,
+            Platform::IbmWatson,
+        ] {
+            let caps = get_supported_ssml(platform);
+            assert!(
+                caps.ssml_elements.iter().any(|e| e.element == "break"),
+                "{:?} missing break",
+                platform
+            );
+        }
+
+        // Eleven v3 does not parse <break> at all; pauses are audio tags.
+        let v3 = get_supported_ssml(Platform::ElevenLabsV3);
+        assert!(v3
+            .ssml_elements
+            .iter()
+            .any(|e| e.element == "[pause] family"));
+        assert!(v3.unsupported.iter().any(|u| u.starts_with("break")));
+    }
+
     #[test]
     fn test_alexa_has_emotion() {
         let caps = get_supported_ssml(Platform::AmazonAlexa);
-        assert!(caps.ssml_elements.iter().any(|e| e.element == "amazon:emotion"));
+        assert!(caps
+            .ssml_elements
+            .iter()
+            .any(|e| e.element == "amazon:emotion"));
     }
 
     #[test]
     fn test_azure_has_express_as() {
         let caps = get_supported_ssml(Platform::MicrosoftAzure);
-        assert!(caps.ssml_elements.iter().any(|e| e.element == "mstts:express-as"));
+        assert!(caps
+            .ssml_elements
+            .iter()
+            .any(|e| e.element == "mstts:express-as"));
     }
 
     #[test]
@@ -506,25 +650,4 @@ mod tests {
         let deserialized: PlatformCapabilities = serde_json::from_str(&json).unwrap();
         assert_eq!(caps, deserialized);
     }
-
-    #[test]
-    fn test_all_platforms_have_break() {
-        for platform in [
-            Platform::AmazonAlexa,
-            Platform::GoogleAssistant,
-            Platform::MicrosoftAzure,
-            Platform::Apple,
-            Platform::W3c,
-            Platform::SamsungBixby,
-            Platform::ElevenLabs,
-            Platform::IbmWatson,
-        ] {
-            let caps = get_supported_ssml(platform);
-            assert!(
-                caps.ssml_elements.iter().any(|e| e.element == "break"),
-                "{:?} missing break",
-                platform
-            );
-        }
-    }
 }
diff --git a/src/formatters/base.rs b/src/formatters/base.rs
@@ -12,6 +12,9 @@ pub enum Platform {
     W3c,
     SamsungBixby,
     ElevenLabs,
+    /// Eleven v3 audio-tag dialect (eleven_v3 / eleven_v3_conversational):
+    /// bracketed natural-language tags instead of SSML.
+    ElevenLabsV3,
     IbmWatson,
 }
 
@@ -26,6 +29,7 @@ impl Platform {
             "w3c" => Some(Platform::W3c),
             "samsung-bixby" | "bixby" => Some(Platform::SamsungBixby),
             "elevenlabs" => Some(Platform::ElevenLabs),
+            "elevenlabs-v3" | "elevenlabs_v3" | "eleven-v3" => Some(Platform::ElevenLabsV3),
             "ibm-watson" | "watson" => Some(Platform::IbmWatson),
             _ => None,
         }
@@ -41,6 +45,7 @@ impl Platform {
             Platform::W3c => "w3c",
             Platform::SamsungBixby => "samsung-bixby",
             Platform::ElevenLabs => "elevenlabs",
+            Platform::ElevenLabsV3 => "elevenlabs-v3",
             Platform::IbmWatson => "ibm-watson",
         }
     }
@@ -105,6 +110,16 @@ pub fn create_formatter(platform: Platform, options: FormatterOptions) -> Box<dy
             // Use the base SSML formatter for W3C standard
             Box::new(super::ssml::SsmlFormatterBase::new(options))
         }
+        Platform::ElevenLabs => {
+            // Pre-v3 prompt markup: <break> tags (and flash-only <phoneme>),
+            // no SSML document, no escaping.
+            Box::new(super::elevenlabs::ElevenLabsFormatter::new(options))
+        }
+        Platform::ElevenLabsV3 => {
+            // Eleven v3 audio-tag dialect: no SSML at all — bracketed
+            // natural-language tags, punctuation pauses, native slash IPA.
+            Box::new(super::elevenlabs_v3::ElevenLabsV3Formatter::new(options))
+        }
         _ => Box::new(super::TextFormatter::new()),
     }
 }