Merge branch 'master' into fix-decoding

CMakeLists.txt

@@ -315,7 +315,8 @@ if (WHISPER_ALL_WARNINGS)
 endif()
 
 if (NOT MSVC)
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Werror=vla")
+    # TODO: temporary disabled until we figure out ggml-metal.m
+    #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Werror=vla")
     #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -ffinite-math-only -funsafe-math-optimizations")
 endif()
 
@@ -505,6 +506,7 @@ else()
 endif()
 
 if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_link_libraries(${TARGET} PUBLIC
         ${CMAKE_DL_LIBS}
         )

Makefile

@@ -117,7 +117,7 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 		CPUINFO_CMD := sysctl machdep.cpu.features machdep.cpu.leaf7_features
 	else ifeq ($(UNAME_S),Linux)
 		CPUINFO_CMD := cat /proc/cpuinfo
-	else ifneq (,$(filter MINGW32_NT% MINGW64_NT%,$(UNAME_S)))
+	else ifneq (,$(filter MINGW32_NT% MINGW64_NT% MSYS_NT%,$(UNAME_S)))
 		CPUINFO_CMD := cat /proc/cpuinfo
 	else ifneq (,$(filter DragonFly FreeBSD,$(UNAME_S)))
 		CPUINFO_CMD := grep Features /var/run/dmesg.boot

bindings/javascript/README.md

@@ -41,7 +41,7 @@ make publish-npm
 
 ## Sample run
 
-```java
+```text
 $ node --experimental-wasm-threads --experimental-wasm-simd ../tests/test-whisper.js
 
 whisper_model_load: loading model from 'whisper.bin'
@@ -63,7 +63,7 @@ whisper_model_load: ggml ctx size =  140.60 MB
 whisper_model_load: memory size   =   22.83 MB
 whisper_model_load: model size    =  140.54 MB
 
-system_info: n_threads = 8 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | NEON = 0 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 1 | BLAS = 0 | 
+system_info: n_threads = 8 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | NEON = 0 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 1 | BLAS = 0 |
 
 operator(): processing 176000 samples, 11.0 sec, 8 threads, 1 processors, lang = en, task = transcribe ...
 

examples/common-ggml.cpp

@@ -64,6 +64,7 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
         case GGML_FTYPE_MOSTLY_IQ2_XXS:
         case GGML_FTYPE_MOSTLY_IQ2_XS:
+        case GGML_FTYPE_MOSTLY_IQ3_XXS:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -195,6 +196,7 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_Q8_K:
                 case GGML_TYPE_IQ2_XXS:
                 case GGML_TYPE_IQ2_XS:
+                case GGML_TYPE_IQ3_XXS:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));

examples/common.cpp

@@ -616,6 +616,7 @@ gpt_vocab::id gpt_sample_top_k_top_p_repeat(
 
 }
 
+
 namespace utf_8 {
     bool is_valid(const std::string &str) {
         uint64_t count = 0; // Count of bytes in the current UTF-8 character
@@ -665,6 +666,21 @@ namespace utf_8 {
     }
 }
 
+bool is_wav_buffer(const std::string buf) {
+    // RIFF ref: https://en.wikipedia.org/wiki/Resource_Interchange_File_Format
+    // WAV ref: https://www.mmsp.ece.mcgill.ca/Documents/AudioFormats/WAVE/WAVE.html
+    if (buf.size() < 12 || buf.substr(0, 4) != "RIFF" || buf.substr(8, 4) != "WAVE") {
+        return false;
+    }
+
+    uint32_t chunk_size = *reinterpret_cast<const uint32_t*>(buf.data() + 4);
+    if (chunk_size + 8 != buf.size()) {
+        return false;
+    }
+
+    return true;
+}
+
 bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
     drwav wav;
     std::vector<uint8_t> wav_data; // used for pipe input from stdin
@@ -689,7 +705,7 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector
 
         fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
     }
-    else if (fname.size() > 256 || fname.size() > 40 && fname.substr(0, 4) == "RIFF" && fname.substr(8, 4) == "WAVE") {
+    else if (is_wav_buffer(fname)) {
         if (drwav_init_memory(&wav, fname.c_str(), fname.size(), nullptr) == false) {
             fprintf(stderr, "error: failed to open WAV file from fname buffer\n");
             return false;

examples/common.h

@@ -140,6 +140,9 @@ namespace utf_8{
 // Audio utils
 //
 
+// Check if a buffer is a WAV audio file
+bool is_wav_buffer(const std::string buf);
+
 // Read WAV audio file and store the PCM data into pcmf32
 // fname can be a buffer of WAV data instead of a filename
 // The sample rate of the audio must be equal to COMMON_SAMPLE_RATE

examples/server/server.cpp

@@ -543,7 +543,76 @@ int main(int argc, char ** argv) {
                              {"Access-Control-Allow-Origin", "*"},
                              {"Access-Control-Allow-Headers", "content-type"}});
 
-    std::string const default_content = "<html>hello</html>";
+    std::string const default_content = R"(
+    <html>
+    <head>
+        <title>Whisper.cpp Server</title>
+        <meta charset="utf-8">
+        <meta name="viewport" content="width=device-width">
+        <style>
+        body {
+            font-family: sans-serif;
+        }
+        form {
+            display: flex;
+            flex-direction: column;
+            align-items: flex-start;
+        }
+        label {
+            margin-bottom: 0.5rem;
+        }
+        input, select {
+            margin-bottom: 1rem;
+        }
+        button {
+            margin-top: 1rem;
+        }
+        </style>
+    </head>
+    <body>
+        <h1>Whisper.cpp Server</h1>
+
+        <h2>/inference</h2>
+        <pre>
+    curl 127.0.0.1:)" + std::to_string(sparams.port) + R"(/inference \
+    -H "Content-Type: multipart/form-data" \
+    -F file="@&lt;file-path&gt;" \
+    -F temperature="0.0" \
+    -F temperature_inc="0.2" \
+    -F response_format="json"
+        </pre>
+
+        <h2>/load</h2>
+        <pre>
+    curl 127.0.0.1:)" + std::to_string(sparams.port) + R"(/load \
+    -H "Content-Type: multipart/form-data" \
+    -F model="&lt;path-to-model-file&gt;"
+        </pre>
+
+        <div>
+            <h2>Try it out</h2>
+            <form action="/inference" method="POST" enctype="multipart/form-data">
+                <label for="file">Choose an audio file:</label>
+                <input type="file" id="file" name="file" accept="audio/*" required><br>
+
+                <label for="temperature">Temperature:</label>
+                <input type="number" id="temperature" name="temperature" value="0.0" step="0.01" placeholder="e.g., 0.0"><br>
+
+                <label for="response_format">Response Format:</label>
+                <select id="response_format" name="response_format">
+                    <option value="verbose_json">Verbose JSON</option>
+                    <option value="json">JSON</option>
+                    <option value="text">Text</option>
+                    <option value="srt">SRT</option>
+                    <option value="vtt">VTT</option>
+                </select><br>
+
+                <button type="submit">Submit</button>
+            </form>
+        </div>
+    </body>
+    </html>
+    )";
 
     // store default params so we can reset after each inference request
     whisper_params default_params = params;
@@ -787,7 +856,13 @@ int main(int argc, char ** argv) {
         } else if (params.response_format == vjson_format) {
             /* try to match openai/whisper's Python format */
             std::string results = output_str(ctx, params, pcmf32s);
-            json jres = json{{"text", results}};
+            json jres = json{
+                {"task", params.translate ? "translate" : "transcribe"},
+                {"language", whisper_lang_str_full(whisper_full_lang_id(ctx))},
+                {"duration", float(pcmf32.size())/WHISPER_SAMPLE_RATE},
+                {"text", results},
+                {"segments", json::array()}
+            };
             const int n_segments = whisper_full_n_segments(ctx);
             for (int i = 0; i < n_segments; ++i)
             {
@@ -801,6 +876,7 @@ int main(int argc, char ** argv) {
                     segment["end"] = whisper_full_get_segment_t1(ctx, i) * 0.01;
                 }
 
+                float total_logprob = 0;
                 const int n_tokens = whisper_full_n_tokens(ctx, i);
                 for (int j = 0; j < n_tokens; ++j) {
                     whisper_token_data token = whisper_full_get_token_data(ctx, i, j);
@@ -815,8 +891,17 @@ int main(int argc, char ** argv) {
                         word["end"] = token.t1 * 0.01;
                     }
                     word["probability"] = token.p;
+                    total_logprob += token.plog;
                     segment["words"].push_back(word);
                 }
+
+                segment["temperature"] = params.temperature;
+                segment["avg_logprob"] = total_logprob / n_tokens;
+
+                // TODO compression_ratio and no_speech_prob are not implemented yet
+                // segment["compression_ratio"] = 0;
+                // segment["no_speech_prob"] = 0;
+
                 jres["segments"].push_back(segment);
             }
             res.set_content(jres.dump(-1, ' ', false, json::error_handler_t::replace),

examples/stream/README.md

@@ -4,7 +4,7 @@ This is a naive example of performing real-time inference on audio from your mic
 The `stream` tool samples the audio every half a second and runs the transcription continously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
-```java
+```bash
 ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
@@ -14,7 +14,7 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 
 Setting the `--step` argument to `0` enables the sliding window mode:
 
-```java
+```bash
  ./stream -m ./models/ggml-small.en.bin -t 6 --step 0 --length 30000 -vth 0.6
 ```
 
@@ -39,8 +39,8 @@ brew install sdl2
 make stream
 ```
 
-Ensure you are at the root of the repo when running `make stream`.  Not within the `examples/stream` dir
-as the libraries needed like `common-sdl.h` are located within `examples`.  Attempting to compile within
+Ensure you are at the root of the repo when running `make stream`. Not within the `examples/stream` dir
+as the libraries needed like `common-sdl.h` are located within `examples`. Attempting to compile within
 `examples/steam` means your compiler cannot find them and it gives an error it cannot find the file.
 
 ```bash