diff --git a/README.md b/README.md
index c24e0951a..446d157fb 100644
--- a/README.md
+++ b/README.md
@@ -653,6 +653,7 @@ The server exposes:
 - `GET /v1/models`
 - `POST /v1/audio/speech`
 - `POST /v1/audio/transcriptions`
+- `POST /v1/audio/alignments`
 - `POST /v1/tasks/run`
 
 More server examples are in [app/server/README.md](app/server/README.md).
diff --git a/app/server/README.md b/app/server/README.md
index 5e6acd0e0..b37030c3c 100644
--- a/app/server/README.md
+++ b/app/server/README.md
@@ -380,6 +380,20 @@ The stream emits `transcript.text.delta` events, one final `transcript.text.done
 
 Note that `stream=true` streams the *output* of an already-uploaded file: the whole recording is sent first, and the deltas describe decoding it. It shortens time-to-first-token on long audio, but nothing can appear while the speaker is still talking. For that, use the live endpoint below.
 
+### `POST /v1/audio/alignments`
+
+Multipart forced-alignment request using uploaded audio bytes and a known transcript. Use this when the server cannot see the client's local audio path, for example when the server is remote or running in Docker.
+
+```bash
+curl http://127.0.0.1:8080/v1/audio/alignments \
+  -F model=qwen3-align \
+  -F language=en \
+  -F text='The task has completed successfully.' \
+  -F file=@/path/to/input.wav
+```
+
+`file`, `model`, and `text` are required; `language` is optional. The selected model must be configured with `task: "align"` and `mode: "offline"`. Uploaded WAV bytes are decoded in memory and are not written to a temporary file. The response includes word timestamps in seconds plus sample offsets.
+
 ### `POST /v1/audio/transcriptions/live`
 
 Streams raw PCM **as it is captured** and returns transcript deltas on the same connection, so partial text can appear while the user is still speaking.
diff --git a/app/server/main.cpp b/app/server/main.cpp
index 8fe29f694..1fec357ef 100644
--- a/app/server/main.cpp
+++ b/app/server/main.cpp
@@ -111,6 +111,8 @@ void print_help() {
         << "       raw PCM in a chunked body, speech audio deltas as SSE on the same connection\n"
         << "  POST /v1/audio/transcriptions\n"
         << "       fields: file, model, language, prompt, stream\n"
+        << "  POST /v1/audio/alignments\n"
+        << "       fields: file, model, text, language\n"
         << "       OpenAI-style streaming: speech stream_format=sse|audio, transcription stream=true\n"
         << "  POST /v1/audio/transcriptions/live?model=<id>\n"
         << "       raw PCM in a chunked body, transcript deltas as SSE on the same connection\n"
diff --git a/app/server/runtime.cpp b/app/server/runtime.cpp
index 74d1e3807..2f2cfbd2d 100644
--- a/app/server/runtime.cpp
+++ b/app/server/runtime.cpp
@@ -795,6 +795,53 @@ std::string task_result_json(const engine::runtime::TaskResult & result, double
     return task_result_json_with_timing(result, timing_json(wall_ms));
 }
 
+std::string alignment_result_json(
+    const engine::runtime::TaskResult & result,
+    const engine::runtime::AudioBuffer & audio,
+    double wall_ms) {
+    if (audio.sample_rate <= 0) {
+        throw std::runtime_error("alignment timing requires a positive input sample rate");
+    }
+
+    std::ostringstream out;
+    out << "{";
+    bool first = true;
+    auto field = [&](const std::string & name) {
+        if (!first) {
+            out << ",";
+        }
+        first = false;
+        out << json_quote(name) << ":";
+    };
+    if (result.text_output.has_value()) {
+        field("text");
+        out << json_quote(result.text_output->text);
+        if (!result.text_output->language.empty()) {
+            field("language");
+            out << json_quote(result.text_output->language);
+        }
+    }
+    field("words");
+    out << "[";
+    for (size_t i = 0; i < result.word_timestamps.size(); ++i) {
+        if (i != 0) {
+            out << ",";
+        }
+        const auto & word = result.word_timestamps[i];
+        out << "{\"word\":" << json_quote(word.word)
+            << ",\"start\":" << (static_cast<double>(word.span.start_sample) / audio.sample_rate)
+            << ",\"end\":" << (static_cast<double>(word.span.end_sample) / audio.sample_rate)
+            << ",\"start_sample\":" << word.span.start_sample
+            << ",\"end_sample\":" << word.span.end_sample
+            << ",\"confidence\":" << word.confidence << "}";
+    }
+    out << "]";
+    field("timing");
+    out << timing_json(wall_ms, audio);
+    out << "}";
+    return out.str();
+}
+
 std::string streaming_task_result_json(
     const engine::runtime::TaskResult & result,
     const std::optional<double> & ttft_ms) {
@@ -1107,6 +1154,9 @@ HttpResponse ServerState::handle(const HttpRequest & request) {
     else if (request.method == "POST" && request.path == "/v1/audio/transcriptions") {
         response = handle_transcription(request);
     }
+    else if (request.method == "POST" && request.path == "/v1/audio/alignments") {
+        response = handle_alignment(request);
+    }
     // Separate path rather than a flag on the endpoint above: the input transport
     // differs (raw chunked PCM vs a complete upload), so keeping it distinct leaves
     // every existing transcription client untouched.
@@ -2588,6 +2638,130 @@ HttpResponse ServerState::run_transcription_stream(
     });
 }
 
+HttpResponse ServerState::handle_alignment(const HttpRequest & request) {
+    std::string content_type;
+    if (const auto it = request.headers.find("content-type"); it != request.headers.end()) {
+        content_type = it->second;
+    }
+    if (const auto boundary = extract_multipart_boundary(content_type)) {
+        return handle_alignment_multipart(request.body, *boundary);
+    }
+    return error_response(
+        400,
+        "audio alignment requests must use multipart/form-data",
+        "invalid_request_error");
+}
+
+HttpResponse ServerState::handle_alignment_multipart(const std::string & body_text, const std::string & boundary) {
+    const auto parts = parse_multipart_body(body_text, boundary);
+    log_multipart_request_summary_if_enabled(config_, parts);
+
+    const MultipartPart * file_part = nullptr;
+    std::string model_id;
+    std::string text;
+    std::string language;
+    std::optional<int> busy_timeout_ms;
+    for (const auto & part : parts) {
+        if (part.name == "file") {
+            file_part = &part;
+        } else if (part.name == "model") {
+            model_id = part.data;
+        } else if (part.name == "text") {
+            text = part.data;
+        } else if (part.name == "language") {
+            language = part.data;
+        } else if (part.name == "busy_timeout_ms") {
+            try {
+                busy_timeout_ms = std::stoi(part.data);
+            } catch (const std::exception &) {
+                return error_response(
+                    400,
+                    "multipart busy_timeout_ms field must be an integer",
+                    "invalid_request_error");
+            }
+            if (*busy_timeout_ms < 0) {
+                return error_response(
+                    400,
+                    "busy_timeout_ms must be >= 0 (0 means no client-side bound)",
+                    "invalid_request_error");
+            }
+        }
+    }
+    if (file_part == nullptr || file_part->data.empty()) {
+        return error_response(
+            400,
+            "multipart alignment request requires a non-empty 'file' field",
+            "invalid_request_error");
+    }
+    if (model_id.empty()) {
+        return error_response(
+            400,
+            "multipart alignment request requires a 'model' field",
+            "invalid_request_error");
+    }
+    if (text.empty()) {
+        return error_response(
+            400,
+            "multipart alignment request requires a non-empty 'text' field",
+            "invalid_request_error");
+    }
+    if (!is_wav_upload_filename(file_part->filename)) {
+        return error_response(
+            400,
+            "only WAV audio uploads are currently supported for alignment",
+            "invalid_request_error");
+    }
+
+    LoadedModel * model_ptr = nullptr;
+    {
+        std::lock_guard<std::mutex> state_lock(models_mutex_);
+        const auto it = model_index_.find(model_id);
+        if (it == model_index_.end()) {
+            return error_response(
+                400,
+                "unknown model id: " + model_id,
+                "invalid_request_error");
+        }
+        model_ptr = models_.at(it->second).get();
+    }
+    auto & model = *model_ptr;
+    if (model.task.task != engine::runtime::VoiceTaskKind::Alignment) {
+        return error_response(
+            400,
+            "audio alignment requires a model configured with task=align",
+            "invalid_request_error");
+    }
+    if (model.task.mode != engine::runtime::RunMode::Offline) {
+        return error_response(
+            400,
+            "audio alignment requires a model configured with mode=offline",
+            "invalid_request_error");
+    }
+
+    engine::runtime::TaskRequest task_request;
+    task_request.audio_input = minitts::cli::read_audio_buffer(std::string_view(file_part->data));
+    task_request.text_input = engine::runtime::Transcript{std::move(text), std::move(language)};
+    task_request = apply_default_request_options(model, std::move(task_request));
+    return run_alignment(model, task_request, busy_timeout_ms);
+}
+
+HttpResponse ServerState::run_alignment(
+    LoadedModel & model,
+    const engine::runtime::TaskRequest & request,
+    std::optional<int> busy_timeout_ms) {
+    if (model_run_mode(model) != engine::runtime::RunMode::Offline) {
+        throw std::runtime_error("audio alignment requires a model configured with mode=offline");
+    }
+    const auto timed_result = run_model(model, request, busy_timeout_ms);
+    if (timed_result.result.word_timestamps.empty()) {
+        throw std::runtime_error("alignment model produced no word timestamps");
+    }
+    if (!request.audio_input.has_value()) {
+        throw std::runtime_error("alignment timing requires audio_input");
+    }
+    return json_response(alignment_result_json(timed_result.result, *request.audio_input, timed_result.wall_ms));
+}
+
 // Live PCM ingest. The client streams raw interleaved samples in a chunked request
 // body while transcript deltas stream back as SSE on the same connection, so
 // partials track capture instead of waiting for a finished upload. Same event shape
diff --git a/app/server/runtime.h b/app/server/runtime.h
index 07c5a52b9..0b7241a4d 100644
--- a/app/server/runtime.h
+++ b/app/server/runtime.h
@@ -178,6 +178,12 @@ class ServerState final : public IHttpHandler {
         LoadedModel & model,
         const engine::runtime::TaskRequest & request,
         std::optional<int> busy_timeout_ms = std::nullopt);
+    HttpResponse handle_alignment(const HttpRequest & request);
+    HttpResponse handle_alignment_multipart(const std::string & body_text, const std::string & boundary);
+    HttpResponse run_alignment(
+        LoadedModel & model,
+        const engine::runtime::TaskRequest & request,
+        std::optional<int> busy_timeout_ms = std::nullopt);
     HttpResponse handle_transcription_live(const HttpRequest & request);
     HttpResponse handle_generic_run(const std::string & body_text);
     HttpResponse handle_generic_stream(const std::string & body_text);