更新

2026-08-03 10:00:05 +08:00
parent 0fb03d0bca
commit 01729b1e0b
109 changed files with 7577 additions and 1153 deletions
@@ -55,6 +55,35 @@ class OpenAIService
        return $model;
    }

+    /**
+     * 语音合成必须使用 OpenAI 兼容协议模型。Dify/ComfyUI 的 API 地址不提供
+     * /audio/speech，因此所选对话模型不兼容时自动回落到已启用的 OpenAI 模型。
+     */
+    public static function getSpeechModel(?int $preferredModelId = null): AiModel
+    {
+        $model = null;
+        if ($preferredModelId) {
+            $model = AiModel::where('id', $preferredModelId)
+                ->where('enabled', 1)
+                ->where('provider', 'openai')
+                ->find();
+        }
+
+        if (!$model) {
+            $model = AiModel::where('enabled', 1)
+                ->where('provider', 'openai')
+                ->order('is_default', 'desc')
+                ->order('sort_order')
+                ->find();
+        }
+
+        if (!$model) {
+            self::throwUnavailableModel('未配置支持语音合成的 OpenAI 兼容模型');
+        }
+
+        return $model;
+    }
+
    public static function getImageModel(?int $preferredModelId = null): AiModel
    {
        $model = null;
@@ -252,6 +281,106 @@ class OpenAIService
        return $data;
    }

+    /**
+     * 使用神经语音模型生成短句 WAV。短句由前端在文本流式输出期间提前提交，
+     * WAV 则避免浏览器额外的解码启动开销。
+     *
+     * @return array{audio: string, content_type: string}
+     */
+    public static function speech(AiModel $model, string $input, string $voice = 'marin'): array
+    {
+        $extra = is_array($model->extra_config ?? null) ? $model->extra_config : [];
+        $ttsModel = trim((string) ($extra['tts_model'] ?? 'gpt-4o-mini-tts'));
+        $ttsVoice = trim((string) ($extra['tts_voice'] ?? $voice));
+        $instructions = trim((string) ($extra['tts_instructions'] ?? (
+            'Speak in natural, warm, conversational Mandarin Chinese. '
+            . 'Use relaxed pacing, subtle emotion, human-like phrasing and short natural pauses. '
+            . 'Avoid an announcer, customer-service, robotic, or overly enthusiastic tone.'
+        )));
+
+        $allowedVoices = [
+            'alloy', 'ash', 'ballad', 'coral', 'echo', 'fable', 'nova',
+            'onyx', 'sage', 'shimmer', 'verse', 'marin', 'cedar',
+        ];
+        if (!in_array($ttsVoice, $allowedVoices, true)) {
+            $ttsVoice = 'marin';
+        }
+        if ($ttsModel === '') {
+            $ttsModel = 'gpt-4o-mini-tts';
+        }
+
+        $legacyTts = in_array($ttsModel, ['tts-1', 'tts-1-hd'], true);
+        if ($legacyTts && !in_array($ttsVoice, ['alloy', 'echo', 'fable', 'onyx', 'nova', 'shimmer'], true)) {
+            $ttsVoice = 'nova';
+        }
+
+        $url = rtrim((string) $model->api_base_url, '/') . '/audio/speech';
+        $payload = [
+            'model'           => $ttsModel,
+            'input'           => $input,
+            'voice'           => $ttsVoice,
+            'response_format' => 'wav',
+        ];
+        if (!$legacyTts && $instructions !== '') {
+            $payload['instructions'] = $instructions;
+        }
+
+        $ch = curl_init($url);
+        curl_setopt_array($ch, [
+            CURLOPT_POST           => true,
+            CURLOPT_POSTFIELDS     => json_encode($payload, JSON_UNESCAPED_UNICODE),
+            CURLOPT_HTTPHEADER     => [
+                'Content-Type: application/json',
+                'Accept: audio/wav',
+                'Authorization: Bearer ' . $model->api_key,
+            ],
+            CURLOPT_RETURNTRANSFER => true,
+            CURLOPT_TIMEOUT        => 30,
+            CURLOPT_CONNECTTIMEOUT => 10,
+            CURLOPT_SSL_VERIFYPEER => false,
+        ]);
+
+        $response = curl_exec($ch);
+        $httpCode = (int) curl_getinfo($ch, CURLINFO_HTTP_CODE);
+        $contentType = (string) (curl_getinfo($ch, CURLINFO_CONTENT_TYPE) ?: 'audio/wav');
+        $curlError = curl_error($ch);
+        curl_close($ch);
+
+        if ($response === false) {
+            throw new HttpResponseException(json([
+                'code'    => 1,
+                'message' => '语音合成连接失败: ' . ($curlError ?: '网络不可达'),
+                'data'    => null,
+            ], 502));
+        }
+
+        if ($httpCode < 200 || $httpCode >= 300) {
+            $detail = self::parseErrorBody($response) ?: ('HTTP ' . $httpCode);
+            throw new HttpResponseException(json([
+                'code'    => 1,
+                'message' => '自然语音生成失败: ' . $detail,
+                'data'    => null,
+            ], 502));
+        }
+
+        if ($response === '') {
+            throw new HttpResponseException(json([
+                'code'    => 1,
+                'message' => '语音服务返回了空音频',
+                'data'    => null,
+            ], 502));
+        }
+
+        if (!str_starts_with(strtolower($contentType), 'audio/')) {
+            $contentType = 'audio/wav';
+        }
+
+        return [
+            'audio'        => $response,
+            'content_type' => $contentType,
+        ];
+    }
+
    /**
     * 测试模型连接是否正常
     * @return array{success: bool, latency_ms: int, reply: string, model: string}