DTMF REJECTION (VoIP)

7.7
During a voice call, DTMF tones are sent to the destination to allow various interactive operations. In PSTN calls, DTMF is not distorted as compared with VoIP. In VoIP, an end-to-end call may use a compression codec such as G729AB, which results in distortion of the DTMF signal. In packet delay or lost situations, DTMF transmission is disturbed and may not necessarily be detected as
a valid digit at the destination. Packet loss creates discontinuity in DTMF transmission even with G.711-based VoIP calls.
The preferred approach in VoIP is to send DTMF as out-of-band packets with redundancy. The digit value, and its power, duration, digit starting, and digit ending information are sent as message packets to the destination. RFC2833 [Schulzrinne and Petrack (2000)] mainly provides guidelines for sending digits as packets. Digits sent this way are called out-of-band packets or out-of-band digits. At the destination, these packets are used to regenerate DTMF digits with right parameters. Before this method can be used, initial negotiation is needed to establish that both VoIP end terminals support out-of-band digit transport through RFC2833.
Additional difficulties with out-of-band support exist. At the source, DTMF detection takes several tens of milliseconds, and during this period, a significant part of the DTMF tones have already transmitted in RTP packets. These uvband tones (inside the IP voice packets) can reach the destination with possible distortion depending on the codec and IP network impediment characteristics. Sometimes this in-band signal as voice is detected as a DTMF digit at destination, which creates the possibility of sending the same digit in in-band and out of band. To avoid this problem, the in- band DTMF signal is removed at the source. In the DTMF specifications, a minimum DTMF tone used for detection is of 23 ms. If the tone transmitted in in-band is comparable with 23 ms, then the destination may start detecting it as a valid digit. Hence, it is required to begin removal of the DTMF tone within 23 ms as the upper limit. Two popular options for DTMF rejection are as follows:
1. Based on a DTMF early warning
2. Delaying all samples at the source to the extent of DTMF normal detection time, usually of greater than 23 ms
I n the previous sections, it is noted that DTMF initial decisions could be arrived with the first 80- to 102-sample data. However, it cannot confirm the digit, but it indicates the likely presence of DTMF tones and can be used to begin DTMF rejection of row and column frequencies.
The removal process may be performed using one of two basic methods. The simplest way is to remove the samples completely and to send silence or comfort noise, but this can be annoying in many situations. A better option would be to pass the signal through row and column frequency notch filters. The filter rejection bandwidth is usually of 20% of the row and column center frequency. It is required to minimize the amount of non-DTMF tone information removal and at the same time ensure that DTMF is removed even if it has drifted by 1.8%. As an example, drifted 1633 Hz can appear anywhere between 1603 and 1663 Hz. Frequencies in between this range have to be removed by the notch filters by 24 to 27 dB. Typically, DTMF tones are generated at -12 to -3 dBm. These tones have to be removed to lower than -27 dBm of the residual DTMF signal. It demands notch filter rejection of about 24 dB
in the frequencies between 1603 and 1663 Hz. If the DTMF is of a lower power (e.g., -12dBm), rejection requirements are of 15 dB to make tone power to below -27 dBm. It is rare, but some phones are generating stronger DTMF tones of the order of 0dBm, and this demands a 27-dB rejection across the band of center frequency with drift.
A fourth-order IIR notch filter with 20% bandwidth can satisfactorily address most rejection requirements. A user may fine-tune the order of filter and bandwidth based on the rejection requirements. A program such as Matlab generates these coefficients with built-in functions for frequency analysis.
Filter rejection output gives the same duration output as the original DTMF tones. Assuming notch filter bandwidth is of 140 Hz, the notch filter introduces an approximately 7-ms delay at the output. In general, DTMF early warning detection also introduces an additional delay of 5 ms. The combined delay is of 12 ms. In most implementations, the combined delay is achieved as 10 to 12 ms to create sufficient margin with an upper allowed limit of 23 ms.
An in- band tone of less than 23 ms creates a tone tick followed by tone discontinuity. It can create a perception of poor quality for the end user or listener on the line. Many deployments insist on removal of initial DTMF tones even though they are not detected as DTMF tones. To make listening comfortable, two main popular options are as follows:
1. Delaying the transmission of the in-band tones by 10 to 20ms. In this scheme, the DTMF algorithm operates on the current samples. These samples are delayed in sending a path by about 20 ms. DTMF detection information is used either to reject tones or to send the original signal. It is the simple way of implementing the DTMF rejection. The main disadvantage of this scheme is of increased end-to-end delay that causes degradation of voice quality.
2. Adding additional DTMF early warning detector on a sample basis helps to reduce the delay. In sample-based early warning, decisions are validated on every sample without waiting for 80- or 102-sample processing. DTMF rejection can start immediately on getting an early warning. Assuming a DTMF early warning is indicated within 1 ms using a TK detector, and that DTMF rejection takes 7 ms to settle, the in-band signal is time limited to 8 ms. This type of early warning is useful to give reduced delay and a better perception of voice quality. The Teager and Kaiser detector can perform such a sample-based tone analysis for early warning of digits.
Figure 7.8 illustrates a DTMF rejection example. The top graph of Fig. 7.8 is a DTMF tone of 50 ms for digit-1. The tone is of 0-dBm power mapped to a |>law calibration of ±8159 sine wave that corresponds to 3.17-dBm power. These tones are passed through cascaded row and column filters as illustrated in Fig. 7.1. The middle graph is for a DTMF rejected signal. The bottom graph is the logarithmic value of rejection power in comparison with the input.

Figure 7.8. DTMF rejection example for digit-1 tones. (a) Digit-1 tone (667 Hz + 1209 Hz) at 0-dBm |i-law power mapping. (b) DTMF rejection output as in-band residue tone. (c) Rejection level in dB, low level clipped to required level of -27 dB.
In Fig. 7.8, when output goes below -27dBm, it is shown clipped to -27dBm for display purposes. By observing the figure, it may be observed that DTMF notch-filter-based rejection is taking approximately 10ms to reach a -27dBm output level for 0-dBm input. The tone power below -27 dBm is not detected as a valid tone by detection algorithms. For DTMF digit input of -10 – dBm power, DTMF rejection takes approximately 5 ms to create 12-dB rejections.