feat(gladia): add full support for Gladia live and batch STT features (#2201)

* feat(gladia): add full support for Gladia live STT features

- Add code switching support: enabled when multiple languages are specified
- Add partial transcripts: receive_partial_transcripts and receive_final_transcripts
- Native multi-channel support already present via channels field
- Custom vocabulary support already present via custom_vocabulary field
- Add custom_metadata field support for metadata passthrough
- Add recommended config for meeting recorders:
  - pre_processing.audio_enhancer: true
  - realtime_processing.words_accurate_timestamps: true

Co-Authored-By: yujonglee <[email protected]>

* fix(gladia): only enable meeting recorder settings for multi-channel mode

The pre_processing.audio_enhancer and realtime_processing.words_accurate_timestamps
settings are designed for multi-channel meeting recorder scenarios. Enabling them
unconditionally for single-channel mode was causing Gladia to not return transcripts.

Now these settings are only enabled when channels > 1.

Co-Authored-By: yujonglee <[email protected]>

* feat(gladia): add diarization_config with name_consistency and custom_vocabulary to batch

- Add diarization_config.name_consistency: true for speaker diarization
- Add custom_vocabulary support from ListenParams.keywords

Co-Authored-By: yujonglee <[email protected]>

* fix

---------

Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: yujonglee <[email protected]>

Author: devin-ai-integration[bot]

owhisper/owhisper-client/src/adapter/argmax/mod.rs

@@ -3,8 +3,6 @@ mod keywords;
 mod language;
 mod live;
 
-pub use batch::{StreamingBatchConfig, StreamingBatchEvent, StreamingBatchStream};
-
 #[derive(Clone, Default)]
 pub struct ArgmaxAdapter;
 

owhisper/owhisper-client/src/adapter/gladia/batch.rs

@@ -36,6 +36,10 @@ struct TranscriptRequest {
     language_config: Option<LanguageConfig>,
     #[serde(skip_serializing_if = "Option::is_none")]
     diarization: Option<bool>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    custom_vocabulary: Option<Vec<String>>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    name_consistency: Option<bool>,
 }
 
 #[derive(Debug, Serialize)]
@@ -191,10 +195,14 @@ impl GladiaAdapter {
             code_switching: (params.languages.len() > 1).then_some(true),
         });
 
+        let custom_vocabulary = (!params.keywords.is_empty()).then(|| params.keywords.clone());
+
         let transcript_request = TranscriptRequest {
             audio_url: upload_result.audio_url,
             language_config,
             diarization: Some(true),
+            custom_vocabulary,
+            name_consistency: Some(true),
         };
 
         let transcript_url = format!("{}/pre-recorded", base_url);

owhisper/owhisper-client/src/adapter/gladia/live.rs

@@ -91,13 +91,20 @@ impl RealtimeSttAdapter for GladiaAdapter {
             let key = api_key.as_deref()?;
             let post_url = Self::build_http_url(&api_base);
 
-            let language_config = (!params.languages.is_empty()).then(|| LanguageConfig {
-                languages: params
-                    .languages
-                    .iter()
-                    .map(|l| l.iso639().code().to_string())
-                    .collect(),
-            });
+            let languages: Vec<String> = params
+                .languages
+                .iter()
+                .map(|l| l.iso639().code().to_string())
+                .collect();
+
+            let language_config = if languages.is_empty() {
+                None
+            } else {
+                Some(LanguageConfig {
+                    code_switching: languages.len() > 1,
+                    languages,
+                })
+            };
 
             let custom_vocabulary = (!params.keywords.is_empty()).then(|| params.keywords.clone());
 
@@ -108,8 +115,16 @@ impl RealtimeSttAdapter for GladiaAdapter {
                 channels,
                 language_config,
                 custom_vocabulary,
+                custom_metadata: None,
                 messages_config: Some(MessagesConfig {
                     receive_partial_transcripts: true,
+                    receive_final_transcripts: true,
+                }),
+                pre_processing: Some(PreProcessing {
+                    audio_enhancer: true,
+                }),
+                realtime_processing: Some(RealtimeProcessing {
+                    words_accurate_timestamps: true,
                 }),
             };
 
@@ -217,17 +232,35 @@ struct GladiaConfig<'a> {
     #[serde(skip_serializing_if = "Option::is_none")]
     custom_vocabulary: Option<Vec<String>>,
     #[serde(skip_serializing_if = "Option::is_none")]
+    custom_metadata: Option<serde_json::Value>,
+    #[serde(skip_serializing_if = "Option::is_none")]
     messages_config: Option<MessagesConfig>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pre_processing: Option<PreProcessing>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    realtime_processing: Option<RealtimeProcessing>,
 }
 
 #[derive(Serialize)]
 struct LanguageConfig {
     languages: Vec<String>,
+    code_switching: bool,
 }
 
 #[derive(Serialize)]
 struct MessagesConfig {
     receive_partial_transcripts: bool,
+    receive_final_transcripts: bool,
+}
+
+#[derive(Serialize)]
+struct PreProcessing {
+    audio_enhancer: bool,
+}
+
+#[derive(Serialize)]
+struct RealtimeProcessing {
+    words_accurate_timestamps: bool,
 }
 
 #[derive(Debug, Deserialize)]
@@ -327,14 +360,6 @@ impl GladiaAdapter {
         let data = msg.data;
         let utterance = data.utterance;
 
-        tracing::debug!(
-            transcript = %utterance.text,
-            is_final = data.is_final,
-            channel = ?utterance.channel,
-            session_id = %session_id,
-            "gladia_transcript_received"
-        );
-
         if utterance.text.is_empty() && utterance.words.is_empty() {
             return vec![];
         }