init

whisper.cpp-1.5.2/examples/talk/.gitignore

@@ -0,0 +1 @@
+audio.mp3

whisper.cpp-1.5.2/examples/talk/CMakeLists.txt

@@ -0,0 +1,8 @@
+if (WHISPER_SDL2)
+    # talk
+    set(TARGET talk)
+    add_executable(${TARGET} talk.cpp gpt-2.cpp)
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
+
+    include(DefaultTargetOptions)
+endif ()

whisper.cpp-1.5.2/examples/talk/README.md

@@ -0,0 +1,41 @@
+# talk
+
+Talk with an Artificial Intelligence in your terminal
+
+[Demo Talk](https://user-images.githubusercontent.com/1991296/206805012-48e71cc2-588d-4745-8798-c1c70ea3b40d.mp4)
+
+Web version: [examples/talk.wasm](/examples/talk.wasm)
+
+## Building
+
+The `talk` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+
+```bash
+# Install SDL2 on Linux
+sudo apt-get install libsdl2-dev
+
+# Install SDL2 on Mac OS
+brew install sdl2
+
+# Build the "talk" executable
+make talk
+
+# Run it
+./talk -p Santa
+```
+
+## GPT-2
+
+To run this, you will need a ggml GPT-2 model: [instructions](https://github.com/ggerganov/ggml/tree/master/examples/gpt-2#downloading-and-converting-the-original-models)
+
+Alternatively, you can simply download the smallest ggml GPT-2 117M model (240 MB) like this:
+
+```
+wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/ggerganov/ggml/resolve/main/ggml-model-gpt-2-117M.bin
+```
+
+## TTS
+
+For best experience, this example needs a TTS tool to convert the generated text responses to voice.
+You can use any TTS engine that you would like - simply edit the [speak](speak) script to your needs.
+By default, it is configured to use MacOS's `say` or `espeak` or Windows SpeechSynthesizer, but you can use whatever you wish.

whisper.cpp-1.5.2/examples/talk/eleven-labs.py

@@ -0,0 +1,20 @@
+import sys
+import importlib.util
+
+if importlib.util.find_spec("elevenlabs") is None:
+    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
+    sys.exit()
+
+from elevenlabs import generate, play, save
+
+# Get a Voice object, by name or UUID
+voice = "Arnold" #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
+
+# Generate the TTS
+audio = generate(
+  text=str(sys.argv[2:]),
+  voice=voice
+)
+
+# Save the TTS to a file
+save(audio, "audio.mp3") 

whisper.cpp-1.5.2/examples/talk/gpt-2.cpp

@@ -0,0 +1,810 @@
+#include "ggml.h"
+#include "common-ggml.h"
+
+#include "gpt-2.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <thread>
+#include <vector>
+#include <regex>
+#include <random>
+
+/////////////////////// GPT-2 BEGIN /////////////////////////
+
+// default hparams (GPT-2 117M)
+struct gpt2_hparams {
+    int32_t n_vocab = 50257;
+    int32_t n_ctx   = 1024;
+    int32_t n_embd  = 768;
+    int32_t n_head  = 12;
+    int32_t n_layer = 12;
+    int32_t ftype   = 1;
+};
+
+struct gpt2_layer {
+    // normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    struct ggml_tensor * ln_2_g;
+    struct ggml_tensor * ln_2_b;
+
+    // attention
+    struct ggml_tensor * c_attn_attn_w;
+    struct ggml_tensor * c_attn_attn_b;
+
+    struct ggml_tensor * c_attn_proj_w;
+    struct ggml_tensor * c_attn_proj_b;
+
+    // mlp
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct gpt2_model {
+    gpt2_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte;     // position embedding
+    struct ggml_tensor * wpe;     //    token embedding
+    struct ggml_tensor * lm_head; // language model head
+
+    std::vector<gpt2_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
+    printf("%s: loading model from '%s'\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != 0x67676d6c) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        fin.read((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != model.hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
+            return false;
+        }
+
+        char word[129];
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+            word[len] = '\0';
+            fin.read((char *) word, len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+
+        ctx_size += (6 + 12*n_layer)*256; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model.wte     = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model.wpe     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
+        model.lm_head = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        // map by name
+        model.tensors["model/ln_f/g"] = model.ln_f_g;
+        model.tensors["model/ln_f/b"] = model.ln_f_b;
+
+        model.tensors["model/wte"]     = model.wte;
+        model.tensors["model/wpe"]     = model.wpe;
+        model.tensors["model/lm_head"] = model.lm_head;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.ln_1_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.ln_2_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, 3*n_embd);
+            layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
+
+            layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
+            layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_mlp_fc_w    = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w  = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
+            layer.c_mlp_proj_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+            model.tensors["model/h" + std::to_string(i) + "/ln_1/g"]        = layer.ln_1_g;
+            model.tensors["model/h" + std::to_string(i) + "/ln_1/b"]        = layer.ln_1_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/ln_2/g"]        = layer.ln_2_g;
+            model.tensors["model/h" + std::to_string(i) + "/ln_2/b"]        = layer.ln_2_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"]    = layer.c_mlp_fc_w;
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"]    = layer.c_mlp_fc_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w;
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int n_mem      = n_layer*n_ctx;
+        const int n_elements = n_embd*n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        size_t total_size = 0;
+
+        bool has_lm_head = false;
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name.data()) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
+                return false;
+            }
+
+            auto tensor = model.tensors[name.data()];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                        __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            // GPT-2 models share the WTE tensor as the LM head
+            if (name == "model/wte" && has_lm_head == false) {
+                memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
+            }
+
+            if (name == "model/lm_head") {
+                has_lm_head = true;
+            }
+
+            total_size += ggml_nbytes(tensor);
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+// TODO: sync latest version from ggml repo
+bool gpt2_eval(
+        const gpt2_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+
+    static size_t buf_size = 512u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    for (int i = 0; i < N; ++i) {
+        ((int32_t *) position->data)[i] = n_past + i;
+    }
+
+    // wte + wpe
+    struct ggml_tensor * inpL =
+        ggml_add(ctx0,
+                ggml_get_rows(ctx0, model.wte, embd),
+                ggml_get_rows(ctx0, model.wpe, position));
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        // norm
+        {
+            // [ 768, N]
+            cur = ggml_norm(ctx0, inpL, 1e-5f);
+
+            // cur = ln_1_g*cur + ln_1_b
+            // [ 768, N]
+            cur = ggml_add(ctx0,
+                    ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                        cur),
+                    ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+        }
+
+        // attn
+        // [2304, 768] - model.layers[il].c_attn_attn_w
+        // [2304,   1] - model.layers[il].c_attn_attn_b
+        // [ 768,   N] - cur (in)
+        // [2304,   N] - cur (out)
+        //
+        // cur = attn_w*cur + attn_b
+        // [2304, N]
+        {
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_attn_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
+                    cur);
+        }
+
+        // self-attention
+        {
+            struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
+            struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
+            struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
+
+            // store key and value to memory
+            if (N >= 1) {
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            // [64, N, 12]
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        ggml_cpy(ctx0,
+                            Qcur,
+                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            // [64, n_past + N, 12]
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3);
+
+            // GG: flash attention
+            //struct ggml_tensor * V =
+            //    ggml_cpy(ctx0,
+            //            ggml_permute(ctx0,
+            //                ggml_reshape_3d(ctx0,
+            //                    ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+            //                    n_embd/n_head, n_head, n_past + N),
+            //                1, 2, 0, 3),
+            //            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
+
+            //struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
+
+            // K * Q
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            // [n_past + N, 64, 12]
+            struct ggml_tensor * V_trans =
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
+
+            // KQV = transpose(V) * KQ_soft_max
+            // [64, N, 12]
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            // [64, 12, N]
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            // [768, N]
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+        }
+
+        // projection
+        // [ 768, 768] - model.layers[il].c_attn_proj_w
+        // [ 768,   1] - model.layers[il].c_attn_proj_b
+        // [ 768,   N] - cur (in)
+        // [ 768,   N] - cur (out)
+        //
+        // cur = proj_w*cur + proj_b
+        // [768, N]
+        {
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
+                    cur);
+        }
+
+        // add the input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        struct ggml_tensor * inpFF = cur;
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = ggml_norm(ctx0, inpFF, 1e-5f);
+
+                // cur = ln_2_g*cur + ln_2_b
+                // [ 768, N]
+                cur = ggml_add(ctx0,
+                        ggml_mul(ctx0,
+                            ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
+                            cur),
+                        ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
+            }
+
+            // fully connected
+            // [3072, 768] - model.layers[il].c_mlp_fc_w
+            // [3072,   1] - model.layers[il].c_mlp_fc_b
+            // [ 768,   N] - cur (in)
+            // [3072,   N] - cur (out)
+            //
+            // cur = fc_w*cur + fc_b
+            // [3072, N]
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_fc_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
+                    cur);
+
+            // GELU activation
+            // [3072, N]
+            cur = ggml_gelu(ctx0, cur);
+
+            // projection
+            // [ 768, 3072] - model.layers[il].c_mlp_proj_w
+            // [ 768,    1] - model.layers[il].c_mlp_proj_b
+            // [3072,    N] - cur (in)
+            // [ 768,    N] - cur (out)
+            //
+            // cur = proj_w*cur + proj_b
+            // [768, N]
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
+                    cur);
+        }
+
+        // input for next layer
+        inpL = ggml_add(ctx0, cur, inpFF);
+    }
+
+    // norm
+    {
+        // [ 768, N]
+        inpL = ggml_norm(ctx0, inpL, 1e-5f);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        // [ 768, N]
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    // inpL = WTE * inpL
+    // [ 768, 50257] - model.lm_head
+    // [ 768, N]     - inpL
+    inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
+
+    // logits -> probs
+    //inpL = ggml_soft_max(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand  (&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result just for the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+/////////////////////////////// GPT-2 END ////////////////////////////////
+
+constexpr int N_THREAD = 8;
+
+struct gpt2_context {
+    std::string prompt_base = R"(Hello, how are you?
+I'm fine, thanks. How are you?
+Thanks, I'm fine too. What are you doing?
+I'm just sitting here.
+It's a lovely day, isn't it?
+Yes, it is. I love the weather this time of year.
+I wish it would rain a little bit.
+Me too.
+)";
+
+    std::mt19937 rng;
+
+    gpt_vocab vocab;
+    gpt2_model model;
+
+    int32_t n_threads = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
+
+    // sampling parameters
+    int32_t top_k = 5;
+    float   top_p = 0.9f;
+    float   temp  = 1.0f;
+};
+
+struct gpt2_context * gpt2_init(const char * path_model) {
+    gpt2_context * ctx = new gpt2_context;
+
+    ctx->rng = std::mt19937(time(nullptr));
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!gpt2_model_load(path_model, ctx->model, ctx->vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, path_model);
+            delete ctx;
+            return nullptr;
+        }
+
+        const int64_t t_load_us = ggml_time_us() - t_start_us;
+
+        printf("gpt-2: model loaded in %d ms\n", (int) (t_load_us/1000));
+    }
+
+    return ctx;
+}
+
+void gpt2_free(struct gpt2_context * ctx) {
+    delete ctx;
+}
+
+const char * gpt2_get_prompt(struct gpt2_context * ctx) {
+    return ctx->prompt_base.c_str();
+}
+
+void gpt2_set_prompt(struct gpt2_context * ctx, const char * prompt) {
+    ctx->prompt_base = prompt;
+}
+
+std::vector<gpt_vocab::id> gpt2_tokenize(const gpt2_context * ctx, const char * text) {
+    return ::gpt_tokenize(ctx->vocab, text);
+}
+
+std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens) {
+    int n_past = 0;
+
+    std::vector<float> embd_w;
+
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt2_tokenize(ctx, text);
+
+    int n_predict = std::min(max_tokens, ctx->model.hparams.n_ctx - (int) embd_inp.size());
+
+    std::vector<gpt_vocab::id> embd = embd_inp;
+
+    size_t mem_per_token = 3000000;
+
+    std::string result;
+
+    for (int i = embd.size(); i < (int) embd_inp.size() + n_predict; i++) {
+        // predict
+        if (!embd.empty()) {
+            if (!gpt2_eval(ctx->model, ctx->n_threads, n_past, embd, embd_w, mem_per_token)) {
+                printf("gpt-2: failed to generate text\n");
+                return "";
+            }
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        {
+            // sample next token
+            const int   top_k = ctx->top_k;
+            const float top_p = ctx->top_p;
+            const float temp  = ctx->temp;
+
+            const int n_vocab = ctx->model.hparams.n_vocab;
+
+            const gpt_vocab::id id = gpt_sample_top_k_top_p(ctx->vocab, embd_w.data() + (embd_w.size() - n_vocab), top_k, top_p, temp, ctx->rng);
+
+            // add it to the context
+            embd.push_back(id);
+        }
+
+        result += ctx->vocab.id_to_token[embd[0]];
+
+        // end of text token
+        if (embd.back() == 50256) {
+            break;
+        }
+    }
+
+    return result;
+}

whisper.cpp-1.5.2/examples/talk/gpt-2.h

@@ -0,0 +1,21 @@
+#pragma once
+
+// TODO: Change to C-style API and move to ./examples for easy reuse.
+
+#include "common.h"
+
+#include <vector>
+#include <map>
+#include <string>
+
+struct gpt2_context;
+
+struct gpt2_context * gpt2_init(const char * path_model);
+void gpt2_free(struct gpt2_context * ctx);
+
+const char * gpt2_get_prompt(struct gpt2_context * ctx);
+void gpt2_set_prompt(struct gpt2_context * ctx, const char * prompt);
+
+std::vector<gpt_vocab::id> gpt2_tokenize(const gpt2_context * ctx, const char * text);
+
+std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens);

whisper.cpp-1.5.2/examples/talk/speak

@@ -0,0 +1,24 @@
+#!/bin/bash
+
+# Usage:
+#  speak.sh <voice_id> <text-to-speak>
+
+# espeak
+# Mac OS: brew install espeak
+# Linux: apt-get install espeak
+#
+#espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"
+
+# Mac OS "say" command
+say "$2"
+
+# Eleven Labs
+# To use it, install the elevenlabs module from pip (pip install elevenlabs)
+# It's possible to use the API for free with limited number of characters. To increase this limit register to https://beta.elevenlabs.io to get an api key and paste it after 'ELEVEN_API_KEY='
+#Keep the line commented to use the free version without api key
+#
+#export ELEVEN_API_KEY=your_api_key
+#wd=$(dirname $0)
+#script=$wd/eleven-labs.py
+#python3 $script $1 "$2"
+#ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3

whisper.cpp-1.5.2/examples/talk/speak.bat

@@ -0,0 +1 @@
+@powershell -ExecutionPolicy Bypass -F examples\talk\speak.ps1 %1 %2

whisper.cpp-1.5.2/examples/talk/speak.ps1

@@ -0,0 +1,12 @@
+# Set-ExecutionPolicy -ExecutionPolicy Bypass -Scope CurrentUser
+param(
+  # voice options are David or Zira
+  [Parameter(Mandatory=$true)][string]$voice,
+  [Parameter(Mandatory=$true)][string]$text
+)
+
+Add-Type -AssemblyName System.Speech;
+$speak = New-Object System.Speech.Synthesis.SpeechSynthesizer;
+$speak.SelectVoice("Microsoft $voice Desktop");
+$speak.Rate="0";
+$speak.Speak($text);

whisper.cpp-1.5.2/examples/talk/talk.cpp

@@ -0,0 +1,375 @@
+// Talk with AI
+//
+
+#include "common-sdl.h"
+#include "common.h"
+#include "whisper.h"
+#include "gpt-2.h"
+
+#include <cassert>
+#include <cstdio>
+#include <fstream>
+#include <regex>
+#include <string>
+#include <thread>
+#include <vector>
+#include <regex>
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t voice_ms   = 10000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
+
+    float vad_thold    = 0.6f;
+    float freq_thold   = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool no_timestamps = true;
+    bool use_gpu       = true;
+
+    std::string person    = "Santa";
+    std::string language  = "en";
+    std::string model_wsp = "models/ggml-base.en.bin";
+    std::string model_gpt = "models/ggml-gpt-2-117M.bin";
+    std::string speak     = "./examples/talk/speak";
+    std::string fname_out;
+};
+
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-vms" || arg == "--voice-ms")      { params.voice_ms      = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
+        else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
+        else if (arg == "-mg"  || arg == "--model-gpt")     { params.model_gpt     = argv[++i]; }
+        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
+        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -vms N,   --voice-ms N    [%-7d] voice duration in milliseconds\n",              params.voice_ms);
+    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
+    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
+    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
+    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -ng,      --no-gpu        [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
+    fprintf(stderr, "  -mg FILE, --model-gpt     [%-7s] gpt model file\n",                              params.model_gpt.c_str());
+    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}
+
+std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+
+    prob = 0.0f;
+    t_ms = 0;
+
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = !params.no_timestamps;
+    wparams.translate        = params.translate;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.max_tokens       = params.max_tokens;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = params.n_threads;
+
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
+
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return "";
+    }
+
+    int prob_n = 0;
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+
+        result += text;
+
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n_tokens; ++j) {
+            const auto token = whisper_full_get_token_data(ctx, i, j);
+
+            prob += token.p;
+            ++prob_n;
+        }
+    }
+
+    if (prob_n > 0) {
+        prob /= prob_n;
+    }
+
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+
+    return result;
+}
+
+const std::string k_prompt =
+R"(This is a dialogue between {0} (A) and a person (B). The dialogue so far is:
+
+B: Hello {0}, how are you?
+A: I'm fine, thank you.
+{1}
+Here is how {0} (A) continues the dialogue:
+
+A:)";
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    // whisper init
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
+
+    // gpt init
+
+    struct gpt2_context * ctx_gpt = gpt2_init(params.model_gpt.c_str());
+
+    // print some info about the processing
+    {
+        fprintf(stderr, "\n");
+        if (!whisper_is_multilingual(ctx_wsp)) {
+            if (params.language != "en" || params.translate) {
+                params.language = "en";
+                params.translate = false;
+                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+            }
+        }
+        fprintf(stderr, "%s: processing, %d threads, lang = %s, task = %s, timestamps = %d ...\n",
+                __func__,
+                params.n_threads,
+                params.language.c_str(),
+                params.translate ? "translate" : "transcribe",
+                params.no_timestamps ? 0 : 1);
+
+        fprintf(stderr, "\n");
+    }
+
+
+    // init audio
+
+    audio_async audio(30*1000);
+    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
+        return 1;
+    }
+
+    audio.resume();
+
+    int n_iter = 0;
+
+    bool is_running  = true;
+    bool force_speak = false;
+
+    float prob0 = 0.0f;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    gpt2_set_prompt(ctx_gpt, "");
+
+    const int voice_id = rand()%6;
+
+    fprintf(stderr, "gpt-2: prompt:\n");
+    fprintf(stderr, "========================\n\n");
+    fprintf(stderr, "%s\n", ::replace(k_prompt, "{0}", params.person).c_str());
+    fprintf(stderr, "========================\n\n");
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        if (!is_running) {
+            break;
+        }
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        int64_t t_ms = 0;
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
+                fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+                audio.get(params.voice_ms, pcmf32_cur);
+
+                std::string text_heard;
+
+                if (!force_speak) {
+                    text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prob0, t_ms));
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\[.*?\\]");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\(.*?\\)");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove all characters, except for letters, numbers, punctuation and ':', '\'', '-', ' '
+                text_heard = std::regex_replace(text_heard, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
+
+                // take first line
+                text_heard = text_heard.substr(0, text_heard.find_first_of('\n'));
+
+                // remove leading and trailing whitespace
+                text_heard = std::regex_replace(text_heard, std::regex("^\\s+"), "");
+                text_heard = std::regex_replace(text_heard, std::regex("\\s+$"), "");
+
+                const std::vector<gpt_vocab::id> tokens = gpt2_tokenize(ctx_gpt, text_heard.c_str());
+
+                if (text_heard.empty() || tokens.empty() || force_speak) {
+                    fprintf(stdout, "%s: Heard nothing, skipping ...\n", __func__);
+                    audio.clear();
+
+                    continue;
+                }
+
+                force_speak = false;
+
+                fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", text_heard.c_str(), "\033[0m", (int) t_ms);
+
+                std::string prompt_base = gpt2_get_prompt(ctx_gpt);
+
+                std::string text_to_speak;
+
+                {
+                    prompt_base += "B: " + text_heard + "\n";
+
+                    std::string prompt = ::replace(::replace(k_prompt, "{0}", params.person), "{1}", prompt_base);
+
+                    text_to_speak = gpt2_gen_text(ctx_gpt, prompt.c_str(), params.max_tokens);
+                    text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
+                    text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of('\n'));
+
+                    // remove first 2 lines of base prompt
+                    if (n_iter > 4) {
+                        {
+                            const size_t pos = prompt_base.find_first_of('\n');
+                            if (pos != std::string::npos) {
+                                prompt_base = prompt_base.substr(pos + 1);
+                            }
+                        }
+                        {
+                            const size_t pos = prompt_base.find_first_of('\n');
+                            if (pos != std::string::npos) {
+                                prompt_base = prompt_base.substr(pos + 1);
+                            }
+                        }
+                    }
+
+                    prompt_base += "A:" + text_to_speak + "\n";
+
+                    {
+                        prompt = ::replace(::replace(k_prompt, "{0}", params.person), "{1}", prompt_base);
+
+                        printf("===============\n");
+                        printf("prompt:\n");
+                        printf("%s\n", prompt.c_str());
+                        printf("===============\n");
+                    }
+                }
+
+                //printf("========================\n");
+                //printf("gpt-2: prompt_base:\n%s\n", prompt_base.c_str());
+                //printf("========================\n");
+
+                gpt2_set_prompt(ctx_gpt, prompt_base.c_str());
+
+                text_to_speak = ::replace(text_to_speak, params.person + ": ", "");
+                int ret = system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
+                if (ret != 0) {
+                    fprintf(stderr, "%s: system() failed!\n", __func__);
+                }
+
+                audio.clear();
+
+                ++n_iter;
+            }
+        }
+    }
+
+    audio.pause();
+
+    whisper_print_timings(ctx_wsp);
+    whisper_free(ctx_wsp);
+
+    return 0;
+}