[resampler] Include the resampler's position in output capture times

When a call stops because the output is full, the next output can be up to one input frame before (frames taken in - taps / 2), which put block times up to about 100 us late when upsampling from 8 kHz. ART's position gives the exact offset.

esphome/components/resampler/microphone/resampler_microphone.cpp

@@ -99,8 +99,8 @@ bool ResamplerMicrophone::init_resampler_(const audio::AudioStreamInfo &input_st
   const audio::AudioStreamInfo output_stream_info(input_stream_info.get_bits_per_sample(),
                                                   input_stream_info.get_channels(), this->target_sample_rate_);
 
-  auto resampler = make_unique<esp_audio_libs::resampler::Resampler>(
-      input_stream_info.ms_to_samples(BUFFER_DURATION_MS), output_stream_info.ms_to_samples(BUFFER_DURATION_MS));
+  auto resampler = make_unique<PositionResampler>(input_stream_info.ms_to_samples(BUFFER_DURATION_MS),
+                                                  output_stream_info.ms_to_samples(BUFFER_DURATION_MS));
 
   esp_audio_libs::resampler::ResamplerConfiguration resample_config = {
       .source_sample_rate = static_cast<float>(input_stream_info.get_sample_rate()),
@@ -165,6 +165,8 @@ void ResamplerMicrophone::process_audio_(const microphone::AudioCapture &capture
 
   const uint8_t *input = capture.data.data();
   uint32_t input_frames = input_stream_info.bytes_to_frames(capture.data.size());
+  // Input frames of this chunk the resampler has taken in so far
+  int64_t frames_consumed = 0;
   while (input_frames > 0) {
     // Stays within the reserved capacity, so this never reallocates
     this->output_buffer_.resize(this->audio_stream_info_.frames_to_bytes(max_output_frames));
@@ -176,10 +178,22 @@ void ResamplerMicrophone::process_audio_(const microphone::AudioCapture &capture
 
     input += results.frames_used * input_bytes_per_frame;
     input_frames -= results.frames_used;
+    frames_consumed += results.frames_used;
 
     if (results.frames_generated > 0) {
+      int64_t output_time_us = 0;
+      if (capture.capture_time_us != 0) {
+        // The resampler holds back half its taps of input, so its next output is near the input frame that far behind
+        // the last one taken in; its position gives the exact offset. The block's first output is one block duration
+        // before that. After a gap in the source, the few frames held back from before it are timed as if there were
+        // no gap.
+        const uint32_t input_rate = input_stream_info.get_sample_rate();
+        output_time_us = capture.capture_time_us + (frames_consumed - this->taps_ / 2) * 1000000 / input_rate +
+                         static_cast<int64_t>(this->resampler_->get_position() * 1000000.0f / input_rate) -
+                         static_cast<int64_t>(results.frames_generated) * 1000000 / this->target_sample_rate_;
+      }
       this->output_buffer_.resize(this->audio_stream_info_.frames_to_bytes(results.frames_generated));
-      this->publish_audio_(this->output_buffer_, this->audio_stream_info_);
+      this->publish_audio_(this->output_buffer_, this->audio_stream_info_, output_time_us);
     } else if (results.frames_used == 0) {
       break;  // No progress; drop the rest of the chunk instead of spinning
     }

esphome/components/resampler/microphone/resampler_microphone.h

@@ -15,6 +15,14 @@
 
 namespace esphome::resampler {
 
+/// @brief The esp-audio-libs resampler, with its exact position for timestamping output.
+class PositionResampler : public esp_audio_libs::resampler::Resampler {
+ public:
+  using Resampler::Resampler;
+  /// @brief Input frames from (last frame taken in - taps / 2) to the next output frame; usually within one frame.
+  float get_position() { return esp_audio_libs::art_resampler::resampleGetPosition(this->resampler_); }
+};
+
 /// @brief Microphone that converts the audio of a source microphone to a different sample rate.
 /// The bits per sample and channels are selected by the source's ``MicrophoneSource``; only the sample rate changes.
 /// Resampling runs in the source microphone's data callback, so it needs no task or ring buffer of its own.
@@ -42,7 +50,7 @@ class ResamplerMicrophone final : public Component, public microphone::Microphon
   void process_audio_(const microphone::AudioCapture &capture);
 
   microphone::MicrophoneSource *source_;
-  std::unique_ptr<esp_audio_libs::resampler::Resampler> resampler_;
+  std::unique_ptr<PositionResampler> resampler_;
   // Reused for every chunk so resampling does not allocate
   std::vector<uint8_t> output_buffer_;