[PATCH 1/2] avcodec/dss_sp, avformat/dss: set the DSS SP sample rate to 11000 Hz

DSS SP synthesises at 12000 Hz and dss_sp_update_state resamples by 11/12,
taking 288 samples per frame down to 264. A frame is therefore 24 ms exactly,
and the output rate is 12000 * 11 / 12 = 11000 Hz. The 264 samples per frame
that this decoder already emits only add up to 24 ms at 11000 Hz; at 11025 a
frame would be 23.9456 ms, which no part of the format accounts for.

Declaring 11025 makes every DSS SP file play 0.23% fast. That is a little over
three seconds of drift across a 23 minute recording, enough to pull a
transcript away from its audio.

The container agrees. A DSS header carries the recording length as an ASCII
HHMMSS field, and dividing the decoded sample count by that length gives the
rate the recorder intended. Over 124 recordings longer than ten minutes, taken
from a mixed set of Olympus and Philips machines, the median implied rate is
11001.6 Hz. The declared length is truncated to the second, so a ten minute
file pins the rate to within 9 Hz; 11025 sits 23 Hz away, outside that margin
on the great majority of the set.

The demuxer sets the packet time base from the same field, so it is corrected
alongside the decoder to keep timestamps and duration consistent.

No decoded sample changes. The fate-dss-sp reference moves by exactly two
lines, the time base and the reported rate; all thirty frame CRCs and all
thirty timestamps are untouched, the latter because the decoder and the
demuxer move together. fate-dss-lp is unaffected, G.723.1 staying at 8000 Hz.
Both tests pass.
---
 libavcodec/dss_sp.c   | 8 +++++++-
 libavformat/dss.c     | 5 ++++-
 tests/ref/fate/dss-sp | 4 ++--
 3 files changed, 13 insertions(+), 4 deletions(-)

diff --git a/libavcodec/dss_sp.c b/libavcodec/dss_sp.c
index 9337371b..a149d5e4 100644
--- a/libavcodec/dss_sp.c
+++ b/libavcodec/dss_sp.c
@@ -291,7 +291,13 @@ static av_cold int dss_sp_decode_init(AVCodecContext *avctx)
 {
     DssSpContext *p = avctx->priv_data;
     avctx->sample_fmt     = AV_SAMPLE_FMT_S16;
-    avctx->sample_rate    = 11025;
+    /* The codec synthesises at 12000 Hz and dss_sp_update_state resamples by
+     * 11/12 -- 288 samples in, 264 out -- so a frame is exactly 24 ms and the
+     * output rate is 12000 * 11 / 12 = 11000 Hz. Declaring 11025 plays every
+     * DSS SP file 0.23% fast, which is a three second drift over a 23 minute
+     * recording. The 264 samples per frame already in this decoder only add up
+     * to 24 ms at 11000 Hz. */
+    avctx->sample_rate    = 11000;
     av_channel_layout_uninit(&avctx->ch_layout);
     avctx->ch_layout      = (AVChannelLayout)AV_CHANNEL_LAYOUT_MONO;
 
diff --git a/libavformat/dss.c b/libavformat/dss.c
index ce86b32d..ea59cb56 100644
--- a/libavformat/dss.c
+++ b/libavformat/dss.c
@@ -146,7 +146,10 @@ static int dss_read_header(AVFormatContext *s)
 
     if (ctx->audio_codec == DSS_ACODEC_DSS_SP) {
         st->codecpar->codec_id    = AV_CODEC_ID_DSS_SP;
-        st->codecpar->sample_rate = 11025;
+        /* See libavcodec/dss_sp.c: the decoder runs at 12000 Hz and
+         * decimates 11:12, so the output rate is 11000 Hz, not 11025. This
+         * also sets the packet time base just below. */
+        st->codecpar->sample_rate = 11000;
         s->bit_rate = 8 * (DSS_FRAME_SIZE - 1) * st->codecpar->sample_rate
                         * 512 / (506 * 264);
     } else if (ctx->audio_codec == DSS_ACODEC_G723_1) {
diff --git a/tests/ref/fate/dss-sp b/tests/ref/fate/dss-sp
index 0984f11e..da7600e7 100644
--- a/tests/ref/fate/dss-sp
+++ b/tests/ref/fate/dss-sp
@@ -1,7 +1,7 @@
-#tb 0: 1/11025
+#tb 0: 1/11000
 #media_type 0: audio
 #codec_id 0: pcm_s16le
-#sample_rate 0: 11025
+#sample_rate 0: 11000
 #channel_layout_name 0: mono
 0,          0,          0,      264,      528, 0xa2579e96
 0,        264,        264,      264,      528, 0xf9b23172
-- 
2.39.5

_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]