Use verbose Whisper transcription output

internal/transcription/client.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"mime/multipart"
@@ -103,6 +104,15 @@ type audioTranscriptionSegment struct {
 	Text  string  `json:"text"`
 }
 
+type audioTranscriptionStatusError struct {
+	status int
+	body   string
+}
+
+func (e audioTranscriptionStatusError) Error() string {
+	return fmt.Sprintf("audio transcription HTTP %d: %s", e.status, e.body)
+}
+
 func New(baseURL string, timeout time.Duration, ffmpegPath string, leadSilence time.Duration) *Client {
 	return NewWithOptions(Options{
 		AudioBaseURL: baseURL,
@@ -233,7 +243,15 @@ func clampTime(v float64) float64 {
 }
 
 func (c *Client) transcribeOpenAIAudio(ctx context.Context, provider ProviderConfig, audio []byte, filename string, in Input) (*audioLLMResponse, time.Duration, error) {
-	return c.doOpenAIAudioTranscription(ctx, provider, audio, filename, in, "json")
+	resp, duration, err := c.doOpenAIAudioTranscription(ctx, provider, audio, filename, in, "verbose_json")
+	if err == nil {
+		return resp, duration, nil
+	}
+	if !isVerboseJSONUnsupported(err) {
+		return nil, duration, err
+	}
+	fallbackResp, fallbackDuration, fallbackErr := c.doOpenAIAudioTranscription(ctx, provider, audio, filename, in, "json")
+	return fallbackResp, duration + fallbackDuration, fallbackErr
 }
 
 func (c *Client) doOpenAIAudioTranscription(ctx context.Context, provider ProviderConfig, audio []byte, filename string, in Input, responseFormat string) (*audioLLMResponse, time.Duration, error) {
@@ -295,7 +313,7 @@ func (c *Client) doOpenAIAudioTranscription(ctx context.Context, provider Provid
 		return nil, duration, fmt.Errorf("audio transcription read: %w", err)
 	}
 	if resp.StatusCode >= 300 {
-		return nil, duration, fmt.Errorf("audio transcription HTTP %d: %s", resp.StatusCode, strings.TrimSpace(string(raw)))
+		return nil, duration, audioTranscriptionStatusError{status: resp.StatusCode, body: strings.TrimSpace(string(raw))}
 	}
 	var out audioTranscriptionResponse
 	if err := json.Unmarshal(raw, &out); err != nil {
@@ -313,6 +331,20 @@ func (c *Client) doOpenAIAudioTranscription(ctx context.Context, provider Provid
 	}, duration, nil
 }
 
+func isVerboseJSONUnsupported(err error) bool {
+	var statusErr audioTranscriptionStatusError
+	if !errors.As(err, &statusErr) {
+		return false
+	}
+	if statusErr.status != http.StatusBadRequest && statusErr.status != http.StatusUnprocessableEntity {
+		return false
+	}
+	body := strings.ToLower(statusErr.body)
+	return strings.Contains(body, "verbose_json") ||
+		strings.Contains(body, "response_format") ||
+		strings.Contains(body, "timestamp_granularities")
+}
+
 func convertAudioSegments(in []audioTranscriptionSegment) []Segment {
 	out := make([]Segment, 0, len(in))
 	for _, s := range in {