diff --git a/docs/public/schemas/llms.json b/docs/public/schemas/llms.json
index 6de1953e38..47f33f8717 100644
--- a/docs/public/schemas/llms.json
+++ b/docs/public/schemas/llms.json
@@ -60,6 +60,10 @@
                         "type": "boolean",
                         "description": "Indicates if speech transcription is supported"
                     },
+                    "speech": {
+                        "type": "boolean",
+                        "description": "Indicates if speech synthesis is supported"
+                    },
                     "openaiCompatibility": {
                         "type": "string",
                         "description": "Uses OpenAI API compatibility layer documentation URL"
diff --git a/packages/core/src/chat.ts b/packages/core/src/chat.ts
index d17bff37a1..8e4e7cb915 100644
--- a/packages/core/src/chat.ts
+++ b/packages/core/src/chat.ts
@@ -152,12 +152,30 @@ export type TranscribeFunction = (
     options: TraceOptions & CancellationOptions
 ) => Promise<TranscriptionResult>
 
+export type CreateSpeechRequest = {
+    input: string
+    model: string
+    voice?: string
+}
+
+export type CreateSpeechResult = {
+    audio: Uint8Array
+    error?: SerializedError
+}
+
+export type SpeechFunction = (
+    req: CreateSpeechRequest,
+    cfg: LanguageModelConfiguration,
+    options: TraceOptions & CancellationOptions
+) => Promise<CreateSpeechResult>
+
 export interface LanguageModel {
     id: string
     completer: ChatCompletionHandler
     listModels?: ListModelsFunction
     pullModel?: PullModelFunction
     transcriber?: TranscribeFunction
+    speaker?: SpeechFunction
 }
 
 async function runToolCalls(
diff --git a/packages/core/src/constants.ts b/packages/core/src/constants.ts
index 26ca93de78..9681a4bfaf 100644
--- a/packages/core/src/constants.ts
+++ b/packages/core/src/constants.ts
@@ -61,6 +61,7 @@ export const SMALL_MODEL_ID = "small"
 export const LARGE_MODEL_ID = "large"
 export const VISION_MODEL_ID = "vision"
 export const TRANSCRIPTION_MODEL_ID = "transcription"
+export const SPEECH_MODEL_ID = "speech"
 export const DEFAULT_FENCE_FORMAT: FenceFormat = "xml"
 export const DEFAULT_TEMPERATURE = 0.8
 export const BUILTIN_PREFIX = "_builtin/"
@@ -200,6 +201,7 @@ export const MODEL_PROVIDERS = Object.freeze<
         bearerToken?: boolean
         listModels?: boolean
         transcribe?: boolean
+        speech?: boolean
         aliases?: Record<string, string>
     }[]
 >(CONFIGURATION_DATA.providers)
diff --git a/packages/core/src/llms.json b/packages/core/src/llms.json
index 65d178c9f8..8b398b45b7 100644
--- a/packages/core/src/llms.json
+++ b/packages/core/src/llms.json
@@ -6,6 +6,7 @@
             "detail": "OpenAI (or compatible)",
             "bearerToken": true,
             "transcribe": true,
+            "speech": true,
             "aliases": {
                 "large": "gpt-4o",
                 "small": "gpt-4o-mini",
@@ -13,7 +14,8 @@
                 "embeddings": "text-embedding-3-small",
                 "reasoning": "o1",
                 "reasoning_small": "o1-mini",
-                "transcription": "whisper-1"
+                "transcription": "whisper-1",
+                "speech": "tts-1"
             }
         },
         {
@@ -21,7 +23,9 @@
             "detail": "Azure OpenAI deployment",
             "listModels": false,
             "bearerToken": false,
-            "prediction": false
+            "prediction": false,
+            "transcribe": true,
+            "speech": true
         },
         {
             "id": "azure_serverless",
diff --git a/packages/core/src/lm.ts b/packages/core/src/lm.ts
index 5045a317b1..aec81336ab 100644
--- a/packages/core/src/lm.ts
+++ b/packages/core/src/lm.ts
@@ -1,4 +1,3 @@
-import { Transform } from "stream"
 import { AICIModel } from "./aici"
 import { AnthropicBedrockModel, AnthropicModel } from "./anthropic"
 import { LanguageModel } from "./chat"
@@ -39,5 +38,6 @@ export function resolveLanguageModel(provider: string): LanguageModel {
     return LocalOpenAICompatibleModel(provider, {
         listModels: features?.listModels !== false,
         transcribe: features?.transcribe,
+        speech: features?.speech,
     })
 }
diff --git a/packages/core/src/openai.ts b/packages/core/src/openai.ts
index 7baf0cd230..1c9b9b1bc7 100644
--- a/packages/core/src/openai.ts
+++ b/packages/core/src/openai.ts
@@ -20,12 +20,13 @@ import {
 import { estimateTokens } from "./tokens"
 import {
     ChatCompletionHandler,
+    CreateSpeechRequest,
+    CreateSpeechResult,
     CreateTranscriptionRequest,
     LanguageModel,
-    PullModelFunction,
 } from "./chat"
 import { RequestError, errorMessage, serializeError } from "./error"
-import { createFetch, iterateBody, traceFetchPost } from "./fetch"
+import { createFetch, traceFetchPost } from "./fetch"
 import { parseModelIdentifier } from "./models"
 import { JSON5TryParse } from "./json5"
 import {
@@ -435,7 +436,6 @@ export async function OpenAITranscribe(
     options: TraceOptions & CancellationOptions
 ): Promise<TranscriptionResult> {
     const { trace } = options || {}
-    const fetch = await createFetch(options)
     try {
         logVerbose(`${cfg.provider}: transcribe with ${cfg.model}`)
         const route = req.translate ? "translations" : "transcriptions"
@@ -453,7 +453,7 @@ export async function OpenAITranscribe(
             method: "POST",
             headers: {
                 ...getConfigHeaders(cfg),
-                ContentType: "multipart/form-data",
+                "Content-Type": "multipart/form-data",
                 Accept: "application/json",
             },
             body: body,
@@ -461,8 +461,10 @@ export async function OpenAITranscribe(
         traceFetchPost(trace, url, freq.headers, freq.body)
         // TODO: switch back to cross-fetch in the future
         const res = await global.fetch(url, freq as any)
+        trace.itemValue(`status`, `${res.status} ${res.statusText}`)
         const j = await res.json()
-        return j
+        if (!res.ok) return { text: undefined, error: j?.error }
+        else return j
     } catch (e) {
         logError(e)
         trace?.error(e)
@@ -470,9 +472,52 @@ export async function OpenAITranscribe(
     }
 }
 
+export async function OpenAISpeech(
+    req: CreateSpeechRequest,
+    cfg: LanguageModelConfiguration,
+    options: TraceOptions & CancellationOptions
+): Promise<CreateSpeechResult> {
+    const { model, input, voice = "alloy", ...rest } = req
+    const { trace } = options || {}
+    const fetch = await createFetch(options)
+    try {
+        logVerbose(`${cfg.provider}: speak with ${cfg.model}`)
+        const url = `${cfg.base}/audio/speech`
+        trace.itemValue(`url`, `[${url}](${url})`)
+        const body = {
+            model,
+            input,
+            voice,
+        }
+        const freq = {
+            method: "POST",
+            headers: {
+                ...getConfigHeaders(cfg),
+                "Content-Type": "application/json",
+            },
+            body: JSON.stringify(body),
+        }
+        traceFetchPost(trace, url, freq.headers, body)
+        // TODO: switch back to cross-fetch in the future
+        const res = await fetch(url, freq as any)
+        trace.itemValue(`status`, `${res.status} ${res.statusText}`)
+        if (!res.ok)
+            return { audio: undefined, error: (await res.json())?.error }
+        const j = await res.arrayBuffer()
+        return { audio: new Uint8Array(j) } satisfies CreateSpeechResult
+    } catch (e) {
+        logError(e)
+        trace?.error(e)
+        return {
+            audio: undefined,
+            error: serializeError(e),
+        } satisfies CreateSpeechResult
+    }
+}
+
 export function LocalOpenAICompatibleModel(
     providerId: string,
-    options: { listModels?: boolean; transcribe?: boolean }
+    options: { listModels?: boolean; transcribe?: boolean; speech?: boolean }
 ) {
     return Object.freeze<LanguageModel>(
         deleteUndefinedValues({
@@ -480,6 +525,7 @@ export function LocalOpenAICompatibleModel(
             id: providerId,
             listModels: options?.listModels ? OpenAIListModels : undefined,
             transcriber: options?.transcribe ? OpenAITranscribe : undefined,
+            speaker: options?.speech ? OpenAISpeech : undefined,
         })
     )
 }
diff --git a/packages/core/src/runpromptcontext.ts b/packages/core/src/runpromptcontext.ts
index f52cc57738..5fc61d1fb4 100644
--- a/packages/core/src/runpromptcontext.ts
+++ b/packages/core/src/runpromptcontext.ts
@@ -30,7 +30,15 @@ import { GenerationOptions } from "./generation"
 import { promptParametersSchemaToJSONSchema } from "./parameters"
 import { consoleLogFormat, stdout } from "./logging"
 import { isGlobMatch } from "./glob"
-import { arrayify, assert, logError, logVerbose, logWarn } from "./util"
+import {
+    arrayify,
+    assert,
+    deleteUndefinedValues,
+    dotGenaiscriptPath,
+    logError,
+    logVerbose,
+    logWarn,
+} from "./util"
 import { renderShellOutput } from "./chatrender"
 import { jinjaRender } from "./jinja"
 import { mustacheRender } from "./mustache"
@@ -39,6 +47,7 @@ import { delay, uniq } from "es-toolkit"
 import {
     addToolDefinitionsMessage,
     appendSystemMessage,
+    CreateSpeechRequest,
     executeChatSession,
     mergeGenerationOptions,
     toChatCompletionUserMessage,
@@ -55,6 +64,7 @@ import {
     DOCS_DEF_FILES_IS_EMPTY_URL,
     TRANSCRIPTION_MEMORY_CACHE_NAME,
     TRANSCRIPTION_MODEL_ID,
+    SPEECH_MODEL_ID,
 } from "./constants"
 import { renderAICI } from "./aici"
 import { resolveSystems, resolveTools } from "./systems"
@@ -82,6 +92,9 @@ import { BufferToBlob } from "./bufferlike"
 import { host } from "./host"
 import { srtVttRender } from "./transcription"
 import { deleteEmptyValues } from "./clone"
+import { hash } from "./crypto"
+import { fileTypeFromBuffer } from "file-type"
+import { writeFile } from "fs"
 
 export function createChatTurnGenerationContext(
     options: GenerationOptions,
@@ -660,15 +673,18 @@ export function createChatGenerationContext(
                 configuration.provider
             )
             if (!transcriber)
-                throw new Error("model driver not found for " + info.model)
+                throw new Error("audio transcribe not found for " + info.model)
             const audioFile = await videoExtractAudio(audio, {
                 trace: transcriptionTrace,
             })
             const file = await BufferToBlob(await host.readFile(audioFile))
             const update: () => Promise<TranscriptionResult> = async () => {
-                trace.itemValue(`model`, configuration.model)
-                trace.itemValue(`file size`, prettyBytes(file.size))
-                trace.itemValue(`file type`, file.type)
+                transcriptionTrace.itemValue(`model`, configuration.model)
+                transcriptionTrace.itemValue(
+                    `file size`,
+                    prettyBytes(file.size)
+                )
+                transcriptionTrace.itemValue(`file type`, file.type)
                 const res = await transcriber(
                     {
                         file,
@@ -703,12 +719,15 @@ export function createChatGenerationContext(
                     update,
                     (res) => !res.error
                 )
-                trace.itemValue(`cache ${hit.cached ? "hit" : "miss"}`, hit.key)
+                transcriptionTrace.itemValue(
+                    `cache ${hit.cached ? "hit" : "miss"}`,
+                    hit.key
+                )
                 res = hit.value
             } else res = await update()
-            trace.fence(res.text, "markdown")
-            if (res.error) trace.error(errorMessage(res.error))
-            if (res.segments) trace.fence(res.segments, "yaml")
+            transcriptionTrace.fence(res.text, "markdown")
+            if (res.error) transcriptionTrace.error(errorMessage(res.error))
+            if (res.segments) transcriptionTrace.fence(res.segments, "yaml")
             return res
         } catch (e) {
             logError(e)
@@ -722,6 +741,71 @@ export function createChatGenerationContext(
         }
     }
 
+    const speak = async (
+        input: string,
+        options?: SpeechOptions
+    ): Promise<SpeechResult> => {
+        const { cache, voice, ...rest } = options || {}
+        const speechTrace = trace.startTraceDetails("🦜 speak")
+        try {
+            const conn: ModelConnectionOptions = {
+                model: options?.model || SPEECH_MODEL_ID,
+            }
+            const { info, configuration } = await resolveModelConnectionInfo(
+                conn,
+                {
+                    trace: speechTrace,
+                    cancellationToken,
+                    token: true,
+                }
+            )
+            if (info.error) throw new Error(info.error)
+            if (!configuration) throw new Error("model configuration not found")
+            checkCancelled(cancellationToken)
+            const { ok } = await runtimeHost.pullModel(configuration, {
+                trace: speechTrace,
+                cancellationToken,
+            })
+            if (!ok) throw new Error(`failed to pull model ${conn}`)
+            checkCancelled(cancellationToken)
+            const { speaker } = await resolveLanguageModel(
+                configuration.provider
+            )
+            if (!speaker)
+                throw new Error("speech converter not found for " + info.model)
+            speechTrace.itemValue(`model`, configuration.model)
+            const req = deleteUndefinedValues({
+                input,
+                model: configuration.model,
+                voice,
+            }) satisfies CreateSpeechRequest
+            const res = await speaker(req, configuration, {
+                trace: speechTrace,
+                cancellationToken,
+            })
+            if (res.error) {
+                speechTrace.error(errorMessage(res.error))
+                return { error: res.error } satisfies SpeechResult
+            }
+            const h = await hash(res.audio, { length: 20 })
+            const { ext } = (await fileTypeFromBuffer(res.audio)) || {}
+            const filename = dotGenaiscriptPath("speech", h + "." + ext)
+            await host.writeFile(filename, res.audio)
+            return {
+                filename,
+            } satisfies SpeechResult
+        } catch (e) {
+            logError(e)
+            speechTrace.error(e)
+            return {
+                filename: undefined,
+                error: serializeError(e),
+            } satisfies SpeechResult
+        } finally {
+            speechTrace.endDetails()
+        }
+    }
+
     const runPrompt = async (
         generator: string | PromptGenerator,
         runOptions?: PromptGeneratorOptions
@@ -968,6 +1052,7 @@ export function createChatGenerationContext(
         prompt,
         runPrompt,
         transcribe,
+        speak,
     })
 
     return ctx
diff --git a/packages/core/src/types/prompt_template.d.ts b/packages/core/src/types/prompt_template.d.ts
index c5445edd19..98724ae767 100644
--- a/packages/core/src/types/prompt_template.d.ts
+++ b/packages/core/src/types/prompt_template.d.ts
@@ -2978,11 +2978,13 @@ type BufferLike =
     | ArrayBuffer
     | ReadableStream
 
+type TranscriptionModelType = OptionsOrString<"openai:whisper-1">
+
 interface TranscriptionOptions {
     /**
      * Model to use for transcription. By default uses the `transcribe` alias.
      */
-    model?: TranscribeModelType
+    model?: TranscriptionModelType
 
     /**
      * Translate to English.
@@ -3054,6 +3056,46 @@ interface TranscriptionResult {
     }[]
 }
 
+type SpeechModelType = OptionsOrString<"openai:tts-1-hd" | "openai:tts-1">
+
+type SpeechVoiceType = OptionsOrString<
+    "alloy",
+    "ash",
+    "coral",
+    "echo",
+    "fable",
+    "onyx",
+    "nova",
+    "sage",
+    "shimmer"
+>
+
+interface SpeechOptions {
+    /**
+     * Speech to text model
+     */
+    model?: SpeechModelType
+    /**
+     * Voice to use (model-specific)
+     */
+    voice?: SpeechVoiceType
+    /**
+     * If true, the transcription will be cached.
+     */
+    cache?: boolean | string
+}
+
+interface SpeechResult {
+    /**
+     * Generate audio-buffer file
+     */
+    filename?: string
+    /**
+     * Error if any
+     */
+    error?: SerializedError
+}
+
 interface ChatGenerationContext extends ChatTurnGenerationContext {
     defSchema(
         name: string,
@@ -3108,6 +3150,7 @@ interface ChatGenerationContext extends ChatTurnGenerationContext {
         audio: string | WorkspaceFile,
         options?: TranscriptionOptions
     ): Promise<TranscriptionResult>
+    speak(text: string, options?: SpeechOptions): Promise<SpeechResult>
 }
 
 interface GenerationOutput {
diff --git a/packages/core/src/types/prompt_type.d.ts b/packages/core/src/types/prompt_type.d.ts
index bf8ce93fc3..f4f7871464 100644
--- a/packages/core/src/types/prompt_type.d.ts
+++ b/packages/core/src/types/prompt_type.d.ts
@@ -336,3 +336,10 @@ declare function transcribe(
     audio: string | WorkspaceFile,
     options?: TranscriptionOptions
 ): Promise<TranscriptionResult>
+
+/**
+ * Converts text to speech.
+ * @param text
+ * @param options
+ */
+declare function speak(text: string, options?: SpeechOptions): Promise<SpeechResult>
diff --git a/packages/sample/genaisrc/speech.genai.mjs b/packages/sample/genaisrc/speech.genai.mjs
new file mode 100644
index 0000000000..7fda27ab13
--- /dev/null
+++ b/packages/sample/genaisrc/speech.genai.mjs
@@ -0,0 +1,4 @@
+
+
+const hello = await speak('Hello, world!');
+console.log(hello)
\ No newline at end of file