@ai-sdk/openai 4.0.108 → 4.0.109

package/src/transcription/openai-transcription-model.ts

@@ -162,6 +162,7 @@ export class OpenAITranscriptionModel implements TranscriptionModelV4 {
     audio,
     mediaType,
     providerOptions,
+    timestampGranularities: requestedTimestampGranularities,
   }: OpenAITranscriptionCallOptions) {
     const warnings: SharedV4Warning[] = [];
 
@@ -194,12 +195,42 @@ export class OpenAITranscriptionModel implements TranscriptionModelV4 {
     const responseFormat =
       openAIOptions?.responseFormat ?? requestConfig.responseFormat;
 
-    // Timestamps are only returned with verbose_json, so only default to
-    // segments when that format is used.
+    // Only verbose_json accepts timestamp_granularities. Diarized JSON
+    // always returns segment timings without a timestamp control.
     const timestampGranularities =
       openAIOptions?.timestampGranularities ??
+      requestedTimestampGranularities ??
       (responseFormat === 'verbose_json' ? ['segment'] : undefined);
 
+    for (const granularity of requestedTimestampGranularities ?? []) {
+      if (
+        openAIOptions?.timestampGranularities != null &&
+        !openAIOptions.timestampGranularities.includes(granularity)
+      ) {
+        warnings.push({
+          type: 'unsupported',
+          feature: 'timestampGranularities',
+          details: `The "${granularity}" timestamp granularity is excluded by providerOptions.openai.timestampGranularities.`,
+        });
+      }
+    }
+
+    for (const granularity of timestampGranularities ?? []) {
+      if (
+        responseFormat !== 'verbose_json' &&
+        !(responseFormat === 'diarized_json' && granularity === 'segment')
+      ) {
+        warnings.push({
+          type: 'unsupported',
+          feature:
+            openAIOptions?.timestampGranularities != null
+              ? 'providerOptions.openai.timestampGranularities'
+              : 'timestampGranularities',
+          details: `OpenAI transcription with the ${responseFormat ?? 'default JSON'} response format does not support ${granularity} timestamps.`,
+        });
+      }
+    }
+
     const chunkingStrategy =
       openAIOptions?.chunkingStrategy ?? requestConfig.chunkingStrategy;
 
@@ -211,7 +242,8 @@ export class OpenAITranscriptionModel implements TranscriptionModelV4 {
       keywords: openAIOptions?.keywords,
       prompt: openAIOptions?.prompt,
       temperature: openAIOptions?.temperature,
-      timestamp_granularities: timestampGranularities,
+      timestamp_granularities:
+        responseFormat === 'verbose_json' ? timestampGranularities : undefined,
     };
 
     for (const [key, value] of Object.entries(transcriptionModelOptions)) {
@@ -314,6 +346,12 @@ export class OpenAITranscriptionModel implements TranscriptionModelV4 {
         : [],
     );
 
+    const words = response.words?.map(word => ({
+      text: word.word,
+      startSecond: word.start,
+      endSecond: word.end,
+    }));
+
     return {
       text: response.text,
       segments:
@@ -322,12 +360,9 @@ export class OpenAITranscriptionModel implements TranscriptionModelV4 {
           startSecond: segment.start,
           endSecond: segment.end,
         })) ??
-        response.words?.map(word => ({
-          text: word.word,
-          startSecond: word.start,
-          endSecond: word.end,
-        })) ??
+        words ??
         [],
+      ...(words != null && words.length > 0 ? { words } : {}),
       language,
       // Duration-billed json responses report the duration only in `usage`:
       durationInSeconds: